Open-Source Humanoid Robotics Landscape (Oct 2026)

Oct 2, 2026 · @julian gilyadov

Summary

As of October 2026, 161 open humanoid-robotics projects are alive — 113 active and 38 slowing — but openness is uneven. Software, simulation and models are rich and mostly permissively licensed; open hardware is small-scale; and the data that matters most (dexterous hands, whole-body motion, real-robot demos) is scarce or non-commercial.

Diagram: open humanoid robotics stack · 6 layers · as of Oct 2, 2026 (shown in the PDF version).

Read across each row: the middle column is what most teams use today; the right column is where open options are thin or legally restricted.

Key findings

• Bodies: Unitree's closed G1 is the de facto research humanoid, and nearly every open controller and humanoid VLA targets it. Open humanoids are lab-scale (Berkeley Humanoid Lite, ToddlerBot) or brand-new kits (Asimov, Roboto Origin), and K-Scale Labs' Nov 2025 shutdown shows how hard open hardware is as a business.
• Hands: ORCA, Aero Hand Open, RUKA-v2 and AmazingHand are the live open hands. They win on modifiability and license, not durability or touch.
• Arms: SO-101 dominates data volume; OpenArm 2.0 and Seeed's reBot are the new research tier with real payload.
• Models: openpi, GR00T and LeRobot are the hubs. Frontier labs stopped opening their best models, 2026's open releases come mostly from Chinese companies and AI2, and GR00T N1.7 is the commercially licensed humanoid option.
• Control: one pipeline — retarget human motion, train a tracker in Isaac Lab or mjlab, deploy on a G1 — limited by motion-data licenses rather than algorithms.
• Simulation: consolidated around Newton/MuJoCo Warp and a modular Isaac Lab 3; value is moving to assets, scenes, calibration and evaluation.
• Data and evaluation: LeRobot v3 is the de facto format, but open data is small, noisy and mostly non-commercial; LIBERO is saturated while real-robot benchmarks score far lower.

The scoreboard below lists every active project; detailed entries follow by category, then recommended stacks and the gaps.

Activity scoreboard

161 open projects showed activity in the past year: 113 are active (commits, releases, new hardware or papers since about April 2026), 38 are slowing (last dated activity 6–12 months ago) and 10 could not be dated. Dormant and discontinued projects are listed under Stalled. Stars are GitHub counts read on Oct 2, 2026 and rounded; each table is sorted by status, then stars.

Humanoids and robot platforms

Hands, grippers, tactile and hand capture

Arms and teleoperation

Actuators and drivers

Robot foundation models

Whole-body control and model-based control

Simulators and environments

Middleware and tooling

Full humanoid platforms (open hardware)

Open humanoid hardware is real but small-scale. The credible 2025–26 designs are lab robots under $6k plus a few full-size kits from Singapore and China. The one venture-backed Western open-humanoid company, K-Scale Labs, shut down in November 2025 (see Stalled).

Berkeley Humanoid Lite

UC Berkeley's printable kid-size humanoid: the cheapest research-grade design whose actuators are open too. Site · paper

• Who: Hybrid Robotics lab under Koushil Sreenath; lead authors Yufeng Chi and Qiayuan Liao (RSS 2025).
• Design: Modular 15:1 cycloidal actuators built on drone motors and printed in PLA on a desktop printer; an Intel N95 mini PC; SteamVR teleop. URDF, MJCF and USD all export from one Onshape model; training runs in Isaac Lab with MuJoCo sim2sim.
• Evidence: Gearboxes run ~90% efficient and track torque within 0.5 N·m. A 60-hour endurance run showed stable efficiency but slowly growing backlash. Walking transfers zero-shot from sim; demos include writing and solving a Rubik's cube.
• Trade-offs: Best for a lab that wants to own every part. Plastic gears wear and payload is low; ToddlerBot is sturdier and cheaper to repair.

ToddlerBot 2.0

Stanford's "ML-first" toddler-size humanoid, built for fast sim-to-real iteration. Site · paper

• Who: Haochen Shi (lead), with Shuran Song and C. Karen Liu (CoRL 2025).
• Design: Dynamixel servos, a Jetson Orin NX and a 200 Hz C++ control loop. Version 2.0 (Aug 2025) added cartwheels and crawling via DeepMimic, Quest VR teleop, stereo depth, MJX training and a rewritten Onshape-to-MJCF pipeline. A Jan 2026 paper extended it to non-foot locomotion.
• Evidence: An independent group rebuilt it from the docs. The maintainers say cartwheel success is "still low" but the robot "almost never breaks".
• Trade-offs: The best-documented small humanoid and native to MuJoCo/MJX. Servo dynamics limit agility, and the design files are non-commercial (CC BY-NC-SA 4.0).

Asimov v1 (Menlo Research)

A full-size, fully open humanoid from Singapore with real QDD actuators. Site

• Who: Menlo Research Pte. Ltd. The legs were open-sourced Jan 27, 2026 and the full robot Apr 27, 2026.
• Design: 120 N·m peak joints, parallel ankles, a Multi Jet Fusion printed structure, Raspberry Pi 5 plus Radxa CM5 compute. The release includes CAD, schematics, PCBs, a MuJoCo model and an mjlab training repo.
• Evidence: The README still lists the locomotion policy, API and app as "pending". The knee survives a ~40 kg quasi-static push (self-reported).
• Trade-offs: The most complete Western full-size open design. The leg actuators alone cost ~$8.5k, no walking policy is released yet, and copyleft licenses (CERN-OHL-S, GPL-2.0) deter some companies.

Roboto Origin (RoboParty)

A Chinese "full-stack" open humanoid kit at roughly half a Unitree G1's price.

• Who: RoboParty, founded Feb 2025 by Yi Huang; $10M seed from MPCi, Xiaomi and Galbot. Open-sourced Jan 2026.
• Design: Damiao QDD actuators, a CNC aluminium frame, D-Robotics RDK X5 compute and a 48 V battery. Isaac Lab RL with an AMP-style gait, ROS 2 Humble and XR teleop. The BOM links every part to a Taobao listing.
• Evidence: Claims ~3 m/s running (self-reported); about 100 kit pre-orders by Feb 2026.
• Trade-offs: The most reproducible supply chain of any open humanoid. The community is Chinese-first (QQ groups), most modules are GPL-3.0, and the actuators are closed.

LeRobot Humanoid (Hugging Face)

Hugging Face's open QDD biped from May 2026, the cheapest legged platform in this list.

• Who: Virgile Batto (built at Hugging Face LeRobot), with Haixuan Xavier Tao of dora-rs.
• Design: Robstride motors (~$1,880 of the BOM), 75 printed parts and a Raspberry Pi 5. The upper body is out of scope for now. Apache-2.0 throughout.
• Notable: It ships a system-identification repo that replays logs in MuJoCo-Warp and fits simulator parameters with CMA-ES — a reusable answer to sim-to-real on cheap hardware.
• Trade-offs: Very new; no walking benchmarks yet.

HopeJR (Hugging Face × The Robot Studio)

A full-size humanoid paired with its own teleop exoskeleton; announced, not yet shipped.

• Who: Rob Knight (The Robot Studio, design) and Martino Russi (Hugging Face, arm, exoskeleton and LeRobot work). Announced May 29, 2025 for delivery "by end of 2025".
• Design: Feetech-servo arms with tendon-driven hands; Robstride series-elastic legs (files uploaded Apr 2026); a no-solder "Homunculus" exoskeleton and glove with a Hall sensor on every joint.
• Evidence: No shipping evidence found as of Oct 2026, but LeRobot already ships hope_jr and homunculus drivers.
• Trade-offs: The only open, low-cost design that pairs a humanoid with a data-collection exoskeleton. Product viability is unproven, and the repo root has no license file (the humanoid folder is Apache-2.0).

OpenLoong / Qinglong

China's state-backed full-size open reference humanoid.

• Who: Shanghai Humanoid Robot Innovation Center, hosted by the OpenAtom Foundation; unveiled at WAIC in Jul 2024.
• Design: Up to 400 TOPS compute, ≥20 kg payload and ≥3 h battery (self-reported). The open files are mostly 2D PDF part drawings, not parametric CAD. MPC+WBC control code and a Unity RL playground are public.
• Trade-offs: A complete full-size reference, but heavy and drawing-based. The Solderpad license is undercut by a "research purposes only" note, and development moved to AtomGit, so the GitHub repos are mirrors.

Tiangong / Tien Kung (X-Humanoid)

Beijing's state-backed humanoid program: open software, closed hardware.

• Who: Beijing Humanoid Robot Innovation Center (X-Humanoid); $100M financing reported Feb 2026; Tiangong 3.0 launched Feb 10, 2026.
• What's open: TienKung-Lab (Isaac Lab locomotion with AMP and GMR retargeting, updated Sep 2026), deployment code, URDF, the RoboMIND dataset, the Pelican-VL model and the XR-1 VLA. No hardware CAD was found despite claims of open "blueprints".
• Evidence: A Tiangong robot won the first humanoid half-marathon (Apr 2025).
• Trade-offs: A good source of locomotion code and data; not buildable.

Fourier N1

A partially open humanoid from Fourier (Shanghai), launched Apr 2025.

• What's open: BOM, structural drawings, assembly guide and base control software (per Yicai); an Apache-2.0 URDF and a MuJoCo Menagerie model. The hardware files' license and location are unverified.
• Evidence: Claims 3.5 m/s running (self-reported). The docs repo was last updated Mar 2026.
• Trade-offs: Good sim models, but thin open tooling compared with Unitree's.

iCub / ergoCub (IIT)

The canonical open research humanoid, maintained by the Italian Institute of Technology for two decades.

• Design: YARP middleware, Gazebo, conda-forge packages and a rich tactile skin; ergoCub is the newer, ergonomics-focused sibling. icub-models 3.2.0 shipped Feb 26, 2026.
• Trade-offs: Deep whole-body-control and tactile heritage. Expensive, built in small numbers, and poorly matched to today's RL and VLA tooling.

Reachy 2 (Pollen Robotics / Hugging Face)

A wheeled, bimanual humanoid upper body with turnkey VR teleop.

• Who: Pollen Robotics (Bordeaux), acquired by Hugging Face in Apr 2025.
• Design: Arms with ~3 kg payload each, an omni base with LiDAR, stereo and RGB-D cameras, ROS 2 Humble and a Python SDK; LeRobot support since Sep 2025.
• Trade-offs: Polished and ready for data collection, but expensive with a small dataset community. The SDK has been quiet since Nov 2025.

Small and expressive robots

Hugging Face's robots with real sales volume are small and expressive, not full humanoids.

• Reachy Mini: a 28 cm desktop robot, $399 tethered or $499 wireless. Over 10,000 units had shipped by late Aug 2026 (secondary source), with 60+ apps on HF Spaces. Software is Apache-2.0; the hardware files are non-commercial.
• Open Duck Mini v2: a ~42 cm, under-$400 bipedal duck inspired by Disney's BDX, led by Antoine Pirrone. Its commercial cousin Microduck ($399, pre-orders from Aug 27, 2026) drew headlines claiming 10,000+ sales within days, which could not be verified. Microduck's software is Apache-2.0 and its gait trains in 1–2 h on one GPU in mjlab (self-reported), but its hardware is closed.
• Upkie: Stéphane Caron's open wheeled biped on mjbots actuators; v13.0 shipped Oct 1, 2026.
• Also active: pib (3D-printed educational humanoid), Alia (tendon-driven, walking only in sim so far) and HumanaOpen (a LeRobot semi-humanoid). InMoov lives on as a hobbyist design.

Closed-hardware humanoids with open software

Unitree's closed robots running Unitree's open software are the de facto research humanoid. Unitree says it shipped 5,500+ humanoids in 2025, mostly to education and research, and nearly every open control or VLA project in this doc supports the G1.

Unitree (G1, R1, H2)

The default body for humanoid research, with the deepest open software stack of any vendor.

• Who: Unitree Robotics (Hangzhou), led by founder Wang Xingxing. It debuted on Shanghai's STAR Market in mid-August 2026, raising ¥6.1B (~$905M), and closed its first day up ~460%. 2025 revenue was reportedly ¥1.70B (+335%).
• Open software: unitree_sdk2 (DDS) and ROS 1/2 drivers; unitree_mujoco; RL training in unitree_rl_gym (Isaac Gym → MuJoCo sim2sim → real) and unitree_rl_lab (Isaac Lab 2.3); xr_teleoperate for Vision Pro, PICO and Quest (v1.6 on Jul 29, 2026 added H2, R1 and BrainCo hands); a LeRobot fork for imitation learning; and two model releases, UnifoLM-WMA-0 (Sep 2025) and UnifoLM-VLA-0 (Jan 2026).
• Adoption: More than 6,500 humanoids produced in 2025 (self-reported). GR00T, BeyondMimic, GMR, HOMIE and most motion-tracking work target the G1, and MuJoCo Menagerie carries its models.
• Trade-offs: Unbeatable price-performance and community, so it is the first embodiment any data or simulation product should support. Actuators and firmware are closed, low-level control reportedly needs EDU units, and a Bluetooth vulnerability ("UniPwn") was disclosed in Sep 2025.

Booster Robotics (T1, K1, T2)

The strongest closed-hardware alternative to Unitree for soccer and kid-size research.

• Who: Booster Robotics (Beijing).
• Open software: booster_gym (Isaac Gym training, MuJoCo cross-sim, Webots checks), now succeeded by booster_train (Isaac Lab), booster_deploy and booster_assets with K1 support. The T1 model is in MuJoCo Menagerie.
• Adoption: T1 was the RoboCup 2025 AdultSize champion platform and K1 the KidSize champion (secondary source).
• Trade-offs: Good hardware for the price, but a far smaller ecosystem than Unitree's.

Fourier (GR-1, GR-2, GR-3)

Shanghai's Fourier publishes deployment code, a LeRobot fork, teleop and MuJoCo models (FFTAI), mostly under LGPL-3.0 or Apache-2.0; the GR-1 URDF is GPL-3.0. The open software is thin next to Unitree's. Its N1 is covered under open hardware above.

AgiBot, Galaxea, EngineAI, HighTorque, LimX, PNDbotics

These Chinese vendors open-source software and robot descriptions, but none has released open humanoid hardware since AgiBot's 2024 X1.

• AgiBot: genie_sim (updated Sep 2026), Genie Envisioner and ACoT-VLA. The X2 humanoid is SDK-only.
• Galaxea: the GalaxeaVLA G0 model family and an Apache-2.0 URDF of the R1 Pro.
• EngineAI: engineai_legged_gym and an Isaac Lab port for its PM01 and T800.
• HighTorque: a BeyondMimic port for the Mini Pi Plus; its only open CAD is the MIT-licensed Panthera-HT arm.
• LimX and PNDbotics: Apache-2.0 and MIT robot descriptions (Oli, TRON, Adam Lite).

ROBOTIS AI Worker

A wheeled, bimanual semi-humanoid (19–25 DoF) built on ROS 2 and LeRobot. The software is Apache-2.0; version 2.2.8 (Sep 18, 2026) added MuJoCo models. The hardware is a closed commercial product.

Hello Robot Stretch

Not a humanoid, but a common open-software mobile manipulator for home research. Stretch 4 sells for $29,950; its body library is LGPL-3.0 and a Stretch 3 model is in MuJoCo Menagerie.

Closed reference points

1X NEO targets homes with a 4-hour battery and remote "Expert Mode" teleoperation. Figure 03, Tesla Optimus, Agility Digit and Apptronik Apollo publish no list prices; Apollo has a Menagerie model. None of their software is open.

Dexterous hands

Four open hands have real 2026 momentum — ORCA, Aero Hand Open, RUKA-v2 and Hugging Face's AmazingHand/HOPEJr — while LEAP remains the most-simulated baseline but has slowed. Open hands now win on hackability and licensing, not price or durability: closed hands claim hundreds of thousands to millions of cycles, open ones report hours.

Prices are charted below; aggregator and paper prices often disagree, so ranges are shown.

Diagram: vendor pages, distributors, aggregators and papers, compiled Oct 2, 2026 · 18 hands (shown in the PDF version).

Open designs cluster under ~$3.5k in parts, but an assembled open hand (ORCA tops out near $7.1k with touch sensing) now overlaps mid-price closed hands such as AgiBot's OmniHand and Inspire's RH56.

LEAP Hand (v1, v2, v2 Advanced)

CMU's Dynamixel hand, still the most-simulated research baseline. Site · v1 paper

• Who: Kenneth Shaw (CMU PhD student, designer) and Deepak Pathak (advisor, Skild AI co-founder), with Ananye Agarwal.
• Design: v1 drives each joint directly with a servo and uses a "universal" knuckle that keeps abduction at any flexion angle; it is ~30% larger than a human hand and assembles in under 4 h. v2 (RSS 2025) is a fully printed hybrid rigid-soft hand; v2 Advanced (Humanoids 2025) adds powered palm joints.
• Evidence: Blind in-hand cube rotation transferred from Isaac Gym. In the v1 paper's cycling test, Allegro "begins to fail after 15 minutes" while LEAP ran the full hour.
• Adoption: The comparison baseline in the ORCA, RUKA and RAPID papers; models in MuJoCo Menagerie, MuJoCo Playground and dex-urdf. No sales or lab counts are published.
• Trade-offs: The easiest hand to model and source. Four fingers, oversized, no touch, a non-standard API license, and no new hardware in 2026 (API last pushed Oct 2025).

ORCA Hand

ETH Zurich's tendon-driven, human-sized hand, now sold by the spin-off ORCA Dexterity. Site · v1 paper · platform paper

• Who: Soft Robotics Lab under Robert Katzschmann; Clemens Christoph and Maximilian Eberlein lead the hardware. ORCA Dexterity, Inc. sells kits and assembled hands; orca_core v0.5.2 shipped Oct 1, 2026.
• Design: One motor per joint with antagonistic nylon tendons, "poppable" joints that dislocate instead of breaking, and automatic tendon calibration. Variants add joint encoders and Hall-effect fingertip force sensing; motors can be Dynamixel or Feetech. LeRobot-native, with MuJoCo models and teleop via Quest 3, MediaPipe or Manus.
• Evidence: Zero-shot sim-to-real ball reorientation after 1 h of Isaac Gym training; learned in-hand reorientation succeeded 9/10 on v2. The v1 tactile skin wore out after 2,000–4,000 cycles.
• Trade-offs: The best open five-finger hand with a living company behind it, MIT-licensed software and CC BY 4.0 hardware files. Tendons need periodic tensioning, and durability is proven in thousands of cycles, not millions.

RUKA and RUKA-v2

NYU's human-sized tendon hand that learns its own controller. Site · v2 paper

• Who: Lerrel Pinto's lab. v1 was led by Anya Zorin and Irmak Guzey; v2 (Mar 2026) adds Raunaq Bhirangi and others.
• Design: Forearm motors pull braided line through PTFE tubes. There are no joint encoders by default; per-finger LSTM controllers are learned from MANUS-glove motion capture. v2 adds a wrist, per-finger abduction, optional magnetic joint encoders and eFlesh fingertips.
• Evidence: v1 completed 29 of 33 GRASP-taxonomy grasps and scored 10/10 on the Kapandji test. v2 cut teleop completion time by 51% and raised success by 21% over v1 across 13 tasks; learned-controller error averages 8.26°.
• Trade-offs: The cheapest human-sized hand with a wrist, and the strongest payload here. Its controllers are trained on MANUS data, which ties good teleop to a $5k+ glove, and tendon tension varies from build to build.

Aero Hand Open (Chestnut Robotics, formerly TetherIA)

The cheapest credible open hand, with a validated tendon simulation model. Paper · docs

• Who: Chestnut Robotics, Inc. (formerly TetherIA); Nan Wang and co-authors (Aug 2026 paper).
• Design: Seven coreless Feetech servos drive Kevlar/Vectran cables, with coupled fingers and a three-way-coupled thumb; an ESP32-S3 controller; about 13 h of printing. Its MuJoCo model, with 20 spatial tendons, was merged into MuJoCo Playground and Menagerie in Nov 2025. Teleop works with Manus, a webcam or Apple Vision Pro.
• Evidence: All 33 GRASP-taxonomy grasps and zero-shot sim-to-real (self-reported).
• Trade-offs: The best price-to-capability ratio and sim fidelity of any open hand. Only seven actuators, modest force, and the CAD is non-commercial (CC BY-NC-SA); release notes show recurring overheat and cable-derailment fixes.

AmazingHand (Pollen Robotics / Hugging Face)

The cheapest and easiest open hand: four fingers on hobby micro-servos.

• Design: Two parallel servos per finger handle flexion and abduction; printed bones with TPU shells; control from Python or an Arduino. Inspired by the ILDA hand.
• Trade-offs: The most-starred open hand repo and permissively licensed (Apache-2.0 code, CC BY 4.0 design). Force, durability and sim models are unpublished, so it suits education and human-robot interaction more than manipulation research.

"DexHand" (three different projects)

• DexHand V1.0 (Rob Knight of The Robot Studio; ROS 2 work by Trent Shumay) has been dormant since 2023–24; the ORCA authors call it hard to assemble.
• BiDexHand (Northwestern, Zhengyang Kris Weng; MIT license) mixes tendon routing and linkages and completes all 33 GRASP grasps; last commit Nov 2025.
• DexHand021 Pro (DexRobot) is a commercial 22-DoF hand; no open hardware was found.

ISyHand and RAPID Hand

• ISyHand (MPI-IS and University of Augsburg, Katherine Kuchenbecker's group; Humanoids 2025) adds an articulated palm, which beat its own fixed-palm version early in RL cube-reorientation training.
• RAPID Hand (Sun Yat-sen, UC Merced and CASIA; NeurIPS 2025) syncs dense fingertip touch and a wrist RGB-D camera within 7 ms. It scored 50/50 on rolling and translation and 24/50 on retrieval. The promised hardware release is unverified.

Closed comparators

• Sharpa Wave sets the 2026 bar for touch and durability: mass production began Oct 2025 and it won a CES 2026 Innovation Award. Its URDF, USD and MJCF models are open even though the hand is not.
• Inspire RH56 is the default hand on Unitree builds and in DexUMI — cheap, but only six actuators.
• Allegro is the legacy research hand; Shadow the legacy high end; Unitree Dex3-1/Dex5-1, Wuji, Robotera XHand, PSYONIC Ability, Tesollo DG-5F and AgiBot OmniHand fill the middle.
• Mid-price closed hands now cost about as much as an assembled open hand, so the case for open hands is modifiability and license, not cost.

Grippers, tactile sensors and hand-data capture

Hand data is the weakest link in the open stack. Gripper capture has scaled to 100,000-trajectory datasets, but open dexterous projects report roughly 40–100 demos per task, Meta's open tactile stack is archived, and most open gloves and exoskeletons are non-commercial or built for a single hand.

Handheld grippers

• UMI (Universal Manipulation Interface; Shuran Song's lab, Columbia then Stanford, with TRI). A GoPro with a fisheye lens and side mirrors rides on a printed gripper, and ORB-SLAM3 recovers the trajectory, so anyone can record demos without a robot. Parts cost $73 plus $298 for the camera. It collected demos 3×+ faster than teleop, reached 71.7% success in unseen environments and transferred across robots 18/20 times. The core repo is mostly static, but the idea lives on in FastUMI, DexUMI, UMI-on-Legs and Trossen's TRumi, and Tsinghua's RDT2 was trained on 10,000+ hours of UMI-style data.
• FastUMI-100K (Shanghai AI Lab). A hardware-decoupled UMI that tracks pose with a RealSense T265; its dataset holds 100k+ trajectories (~600 h) across 54 task categories, and π0 fine-tunes reach 73–100% on short tasks.
• PincOpen (Pollen Robotics). A ~€25 printable parallel gripper for the SO-ARM100, derived from Reachy 2's gripper; CC BY-SA 4.0.
• SO-101 and ALOHA jaws are the default grippers behind most LeRobot data; ALOHA's original repo is dormant while Trossen sells the continuation.

Tactile sensors

Tactile simulation is early. Isaac Lab's TacSL sensor is marked experimental and modeled on a discontinued GelSight prototype. Taccel (Peking University, BIGAI and UCLA; NeurIPS 2025) runs peg insertion at 915 FPS across 4,096 environments on one H100 (~18× real time), but its code release is unverified. TACTO and DiffTactile are dormant.

Gloves, exoskeletons and retargeting

• Evidence: DexUMI was 3.2× more efficient than teleop with 86% average task success, but each new hand needs its own exoskeleton (CoRL 2025 best-paper finalist). DOGlove's force feedback raised in-hand rotation to 10/10, versus 1/10 for vision-only AnyTeleop.
• Retargeting: dex-retargeting (UCSD/NVIDIA, MIT) still sees ~121k PyPI downloads a month but has been idle since Aug 2025; Meta's GeoRT trains in 1–2 minutes but is non-commercial; dex-urdf lacks ORCA, RUKA, Aero and XHand models.
• Gap: no open, commercially licensable glove records both kinematics and force and retargets to many hands — the commercial options cost $1.5k–$7.3k per glove, and the open ones are research-grade or license-restricted.

Arms and bimanual rigs

Open arms have split into two tiers. The SO-101 family dominates data volume — 9,400+ Hugging Face datasets carry "so101" in their name — while 2026's research-grade data is moving to $1.5k–$7k quasi-direct-drive (QDD) arms such as OpenArm 2.0 and Seeed's reBot, which still have very few public datasets.

SO-100 / SO-101 (The Robot Studio × Hugging Face)

The default first arm for robot learning, and the center of gravity of the LeRobot community.

• Who: The Robot Studio (France) designs the hardware; Hugging Face's LeRobot team maintains the software. HF CEO Clem Delangue called SO-101 "the first robot arms any AI builder should buy". About nine vendors sell kits, including Seeed, WowRobo and PartaBot.
• Design: Six bus servos per arm over USB; the leader arm uses mixed gear ratios so it can be back-driven by hand. It has MuJoCo Menagerie models, Isaac Lab support through LeIsaac, and a browser GUI (LeLab).
• Adoption: No official sales figure exists. Beyond the name count above, 4,033 datasets mention SO-100, SO-100s make up 31.4% of LeRobot's curated community dataset, and phospho reports 2,000+ robots on its app (self-reported).
• Evidence: Community-data pretraining lifted SmolVLA's real SO-100 success from 51.7% to 78.3%. ACT trains on ~50 episodes in 30–60 minutes on an RTX 4090.
• Trade-offs: The cheapest way into the biggest ecosystem. Hobby servos mean low payload, backlash and stripped gears (the calibration tool caps torque at 30%), and the sim files ship two calibration conventions.

OpenArm 2.0 (Enactic)

The only fully open, human-scale 7-DoF arm with real payload, now a complete data-collection system.

• Who: Enactic, Inc. (Tokyo), led by CEO Hiroto Yamamoto; built by WowRobo, Anvil Robotics, Cereboto and others.
• Design: Back-drivable QDD joints over CAN-FD at 1 kHz (Linux only). Version 2.0 (2026) adds a gripper with an in-hand camera. The KER leader is motorless and worn like a backpack (1.7 kg, 70%-scale links, 1:1 joint mapping, $2.4k–$2.6k). The OpenArm Cell is a standardized evaluation enclosure built so that "Model A outperforms Model B" means the same thing across labs. Its own dataset format converts to LeRobot and GR00T.
• Adoption: First-class LeRobot support, but only 39 Hugging Face datasets are tagged openarm.
• Trade-offs: Bilateral, force-rich teleop that SO-class arms can't do. Linux-only CAN, dependence on Damiao motors, CERN-OHL-S reciprocity and a bespoke data format are the costs.

Seeed reBot Arm B601

Seeed's open "step-up" arm, launched Apr 2026 and already 3.9k stars.

• Design: Sheet metal and printed parts with a BOM down to the screw; Damiao (24 V) or RobStride (48 V) variants. Isaac Sim USD files with an NVIDIA sim-to-real course, a browser MuJoCo twin, a Menagerie model and LeRobot core support.
• License: OSHWA-certified; moved from a non-commercial license to CERN-OHL-W in May 2026 — a sign vendors are moving toward commercial-friendly terms.
• Trade-offs: Fills the gap between SO-101 and OpenArm. Six joints, not seven, and its servo leader cannot do bilateral teleop.

Other open arms

• Koch v1.1 and Moss (Jess Moss, after Alexander Koch's original): Dynamixel and Feetech predecessors of SO-100. Koch is still supported in LeRobot but costs about twice as much; Moss is deprecated.
• PAROL6 (Petar Crnjak / Source Robotics; GPL-3.0): an industrial-style printed stepper arm with a GUI. It is open-loop and has no brakes, so it "falls on power loss", and its docs warn long sessions can overheat the PETG structure — a poor fit for imitation-learning data.
• AR4 (Annin Robotics; custom license): machined-aluminium stepper arm sold as kits; an MK4 manual appeared Feb 2026.
• Arctos and Thor: low-cost printed stepper arms for makers; Arctos's ROS stack targets ROS Melodic, and Thor looks dormant.
• ROBOTIS OMX and OMY: Dynamixel arms that ship factory-calibrated (no calibration step) and are in LeRobot core; OMY adds cobot-style gravity compensation.
• Waveshare RoArm-M3: an ESP32 Wi-Fi arm with a web UI; very low payload.
• AM-ARM200 and AlohaMini (Li Yiteng, 2026): an SO-class upgrade with a 1 kg payload, and a ~$950 dual-arm mobile robot with a lift (1.1k stars).

Bimanual and mobile rigs

• ALOHA / ALOHA 2 (Tony Zhao and Chelsea Finn; Stanford and Google DeepMind): the recipe that made low-cost bimanual imitation learning work. With 50 demos, ACT scored 84–96% on three of four reported tasks (64% on tape). The original repo is stuck on ROS 1; Trossen's kits and SDKs carry it forward, and 155 Hub datasets are tagged aloha.
• XLeRobot (Gaotian "Vector" Wang, Rice University): two SO-101 arms on a LeKiwi base and an IKEA cart, ~$660 in parts or a $579 kit. Its catalog lists 43 community uses, including papers at ACL, AAAI and HRI 2026.
• LeKiwi (SIGRobotics UIUC × Hugging Face): the omni-wheel base for SO-101, built into LeRobot.
• Closed arms that open stacks support: I2RT YAM (GELLO's recommended arm and the source of AI2's 720-hour MolmoAct2 dataset), AgileX Piper, Trossen WidowX, Franka FR3, Kinova Gen3, UR5e and Unitree Z1 all have LeRobot plugins or Menagerie models.

Teleoperation and data-collection systems

Teleop software is commoditizing quickly, much of it driven by NVIDIA, Hugging Face and robot vendors. What stays hard is throughput, operator fatigue, force feedback and data quality — not the software that streams poses.

• GELLO (Philipp Wu, Pieter Abbeel et al., UC Berkeley) remains the reference for cheap joint-space teleop; its user study found it more reliable than VR controllers or a SpaceMouse. Franka now maintains its FR3 integration, and Wu went on to co-found data company XDOF. The CAD is available only by email request.
• NVIDIA Isaac Teleop pitches itself as "the unified framework for high-fidelity ego-centric and robotics data collection" and already records human egocentric data without a robot — directly competitive with data-capture startups.
• TWIST2 (Stanford and Amazon FAR) is the cheapest complete open loop from humanoid teleop to policy: 100 demos in 15 minutes at near-100% success (self-reported), though PICO body pose is weak at elbows and knees.
• LeRobot v0.6 added DAgger rollouts (take over with the leader arm or a foot pedal; interventions are tagged), phone teleop and an Isaac Teleop bridge.
• phosphobot (Paris, YC) also sells custom datasets and models, but releases stopped in Oct 2025 as LeRobot and LeLab absorbed its role.
• Dormant: Open-TeleVision (its lineage lives on in Unitree's televuer), Bunny-VisionPro, ACE and AnyTeleop (only dex-retargeting is open).
• Pain points: No VR or phone system gives force feedback; Vision Pro streaming adds 50–100 ms; powered leaders tire operators over long sessions (the reason KER exists); and each system writes its own format.

Actuators and motor drivers

Open actuators remain niche. Nearly every open humanoid runs on closed Chinese QDD motors (Damiao, Robstride, Encos) or hobby servos, and the one fully open, high-performance motor controller with momentum is mjbots' moteus.

Closed parts the open projects actually buy (prices from project BOMs):

• Feetech STS3215 — $13.89 each, 16.5 kg·cm (6 V) or 30 kg·cm (12 V version); inside SO-101, XLeRobot, HopeJR and HumanaOpen.
• Dynamixel — ~$180 per motor in ToddlerBot's 24-motor bundle; precise but expensive.
• Damiao — DM-4340P at ¥949 and DM-10010L at ¥1,989 in Roboto Origin's BOM; also OpenArm's and reBot's motors.
• Robstride — ~$157 per motor in LeRobot Humanoid's BOM; RS02 in HopeJR's legs.
• Encos — Asimov's leg actuators and joint parts total ~$8.5k for 12 DoF.
• Unitree joint motors are sold only inside Unitree robots.

Why it matters: cheap actuators are where sim-to-real breaks. Every low-cost humanoid project builds its own system-identification rig (ToddlerBot's motor test bed, LeRobot Humanoid's CMA-ES fitting, Microduck's backlash and voltage-sag randomization). No shared, validated actuator-model library exists for the popular motors, even though MuJoCo 3.5 added a system-identification toolbox and actuator delays.

Robot foundation models (open-weight VLAs and world models)

Three hubs dominate open robot models: Physical Intelligence's openpi, NVIDIA's GR00T and Hugging Face's LeRobot, which now wraps most of the others. But the frontier labs stopped opening their best models in 2025–26, most new open releases come from Chinese companies and AI2, and saturated simulation benchmarks hide much weaker real-robot performance.

Almost all benchmark numbers are self-reported. Hub download counts are rolling file requests and differ between model cards and listings by up to 30×, so treat them as order-of-magnitude.

Physical Intelligence openpi (π0, π0-FAST, π0.5)

The reference open generalist VLA family, with the most real-world evidence. GitHub · π0.5 paper

• Who: Physical Intelligence (San Francisco), founded by Karol Hausman, Sergey Levine, Chelsea Finn, Brian Ichter, Lachy Groom, Adnan Esmail and Quan Vuong. It raised $400M in 2024 ($2.4B valuation), $600M in 2025 ($5.6B), and reportedly ~$1B in 2026 at ~$11B.
• Design: A PaliGemma vision-language model plus a flow-matching "action expert" running up to 50 Hz. π0.5 co-trains on web data, subtask commands, cross-embodiment data and ~400 h of mobile manipulation to generalize to new homes.
• Adoption: 633 Hub fine-tunes of π0.5; it is the baseline in BEHAVIOR and RoboChallenge, and the BEHAVIOR 2025 winner was built on it.
• Robustness: On the independent LIBERO-Plus benchmark, π0 falls to 15.8% under camera changes and 6.6% under new robot start poses. LIBERO-PRO found π0 and π0.5 near 0% when objects move.
• Trade-offs: The strongest open generalist and largest community. It is JAX-first (users report PyTorch/JAX accuracy gaps), fine-tuning wants an 80 GB GPU, and no new open checkpoints have appeared since Sep 2025: π*0.6 (Nov 2025) and PI's memory work are closed, and users are asking whether more will be opened (issue #980).

NVIDIA Isaac GR00T (N1.5 → N1.7, N2 preview)

The open VLA line built around humanoids and hands, now commercially licensed. GitHub · model card

• Who: NVIDIA's GEAR lab (Jim Fan, Yuke Zhu) and the Isaac team.
• Design: N1.7 pairs Cosmos-Reason2-2B (Qwen3-VL architecture) with a flow-matching DiT. Pretraining used 20,000 h of "EgoScale" human video plus robot data, in a relative end-effector action space shared by humans and robots. It reaches finger-level control and drives humanoids (Unitree G1, AgiBot G1) through the SONIC whole-body controller.
• License shift: N1.5 and N1.6 weights were non-commercial; NVIDIA staff called it "a data issue, not a license choice". N1.7 fixed that in Apr 2026.
• Compute: 27.9 ms per inference on an H100 with TensorRT (~36 Hz) but 216.5 ms on Jetson Orin (~4.6 Hz); fine-tuning wants 40 GB+ GPUs.
• Adoption: NVIDIA names AGIBOT, LG Electronics and NEURA Robotics among N1.7 users (self-reported). GR00T N2, a world-action model, was previewed Mar 16, 2026 for release by end of 2026.
• Trade-offs: The best open choice for humanoids, hands and commercial terms, with NVIDIA's toolchain around it. Three versions in a year, config confusion in the issue tracker, and slow inference on Orin.

SmolVLA (Hugging Face)

The cheapest open VLA to run and adapt, with by far the most fine-tunes (7,808). Model · paper

• Design: A SmolVLM2 backbone with a ~100M flow-matching expert, pretrained on 481 community datasets (22.9k episodes, mostly SO-100 arms). Asynchronous inference cuts response time ~30%.
• Practicalities: 20k training steps take ~4 h on one A100. The docs say to collect at least 50 episodes: "25 episodes… was not enough".
• Trade-offs: Runs on consumer GPUs and lives inside LeRobot. Lower accuracy ceiling, and its pretraining is skewed toward cheap arms, not humanoids or hands.

OpenVLA and OpenVLA-OFT (Stanford)

• OpenVLA (Jun 2024): DINOv2 and SigLIP vision encoders with Llama 2, trained on 970k Open X-Embodiment episodes. Still the most-downloaded robotics model on the Hub, but superseded.
• OpenVLA-OFT (Moo Jin Kim, Chelsea Finn, Percy Liang; Feb 2025): a fine-tuning recipe — parallel decoding, action chunks, continuous actions with L1 loss — giving 26× throughput. It is the most robust model on LIBERO-Plus (59.7% under camera shifts).
• Trade-offs: The standard single-arm baseline, but 7B, trained on 2024-era data and bound by Llama 2 terms.

X-VLA (Tsinghua AIR + Shanghai AI Lab)

A 0.9B Florence-2-based model with per-embodiment "soft prompts", pretrained on 290k episodes across 7 platforms. It won the IROS 2025 AgiBot World Challenge and posts the best simulation scores per parameter (CALVIN 4.43), but its real-robot results are weak (chart below).

RDT2 (Tsinghua)

A Qwen2.5-VL-7B model trained on 10,000+ hours of human demos captured with an improved handheld UMI gripper across 100+ indoor scenes. It claims zero-shot deployment on unseen robots (bimanual UR5e, Franka FR3) and was accepted at ICML 2026. It is the clearest open evidence that wearable or handheld human data scales into robot policies — though only for grippers.

MolmoAct2 (AI2)

The most fully open 2026 release: weights, training code and all training data under Apache-2.0. Model

• Design: 5B, built on Molmo2-ER (an embodied-reasoning model) with a flow-matching expert and OpenFAST, an open re-implementation of the FAST tokenizer. Its data includes 720 h of bimanual YAM teleop, which AI2 calls the largest open bimanual dataset.
• Evidence: AI2 reports beating π0.5 on seven benchmarks (self-reported) and a third-party Cortex AI evaluation placing it first on 7 of 8 tasks.
• Trade-offs: The best base when provenance matters (US nonprofit, data included). Independent replications are still scarce.

The 2026 Chinese wave

Most 2026 open VLAs share one recipe: a 3–6B model with a Qwen-VL backbone and a flow-matching DiT action expert.

• Xiaomi-Robotics-0 (Feb 2026): top self-reported sim scores (CALVIN 4.75–4.80) and 80 ms inference on an RTX 4090.
• Spirit v1.5 (Spirit AI, Jan 2026): ranked #1 on RoboChallenge at release; trained "largely on open-ended, goal-driven diverse data" instead of scripted demos — evidence that unscripted collection pays.
• GigaBrain-0.7 (GigaAI, Aug 2026): a "three-system" model trained on 37,000+ hours of mixed embodied data.
• DM0 / DM0.5 (Dexmal, Jul 2026): parameters undisclosed; 93.6 on RoboTwin 2.0 (self-reported). Dexmal co-runs RoboChallenge with Hugging Face, a conflict of interest for its leading scores there.
• LingBot-VLA (Robbyant, Ant Group, Jan 2026): trained on 20,000 h of real data from 9 dual-arm setups, yet scores only ~16–17% on its own GM-100 benchmark (which it says beats π0.5 and GR00T N1.6).
• WALL-OSS (X Square Robot): version 0.5 shipped May 2026; the company reportedly raised a $140M Series A++.
• Galaxea G0.5 (Jun 2026): built for wheeled mobile manipulators; Galaxea switched to a non-commercial license from G0.5 on.

Shanghai AI Lab and AgiBot

• InternVLA-M1 (MIT) adds spatial grounding before action training. InternVLA-A1 (non-commercial) mixes understanding, future-frame generation and action experts, trained on 692M+ frames of real, synthetic and human-video data — a strong signal for synthetic-data vendors.
• EO-1 (MIT) interleaves reasoning and action in one decoder; its repo moved in Nov 2025.
• AgiBot GO-1 (Sep 2025) is trained on AgiBot World's 1M trajectories but is non-commercial; GO-2 (Apr 2026) is closed.

Models aimed at humanoids and hands

• Unitree UnifoLM-VLA-0 (Jan 2026): a G1-focused VLA trained on 12 open datasets plus G1 data.
• Being-H0 (BeingBeyond; ICML 2026): pretrained on human hand-motion video with MANO hand models, one of the few open models built for dexterous hands.
• GR00T N1.7 (above) is the only major open model with finger-level humanoid control.

World models and world-action models

The 2026 trend is video-generation backbones that predict future frames and actions together. They need large action-labelled video corpora and are heavy to run.

• NVIDIA Cosmos 3 (May 2026): a reasoner-plus-generator "omni-model" in Edge, Nano and Super sizes, commercially usable. Its 16B Nano-Policy-DROID variant claims #1 on RoboArena (self-reported).
• DreamZero (NVIDIA GEAR, Feb 2026): 14B on the Wan2.1 video model, ~7 Hz closed loop, non-commercial; the precursor of GR00T N2.
• Genie Envisioner (AgiBot): a video world model, action decoder and neural simulator; GE-Act v2.0 shipped Sep 10, 2026 (non-commercial).
• UnifoLM-WMA-0 (Unitree, Sep 2025): doubles as a synthetic-data engine and a policy enhancer.
• V-JEPA 2 / 2.1 (Meta): self-supervised video encoders; the action-conditioned variant plans zero-shot on a Franka (100% reaching, 50–80% pick-and-place) but takes ~16 s per action.
• OpenWAM-Alpha (Sep 2026): publishes its pretraining mix — 70% robot data (40% real, 30% simulated) and 30% egocentric human video.

Classic single-task policies

Diffusion Policy (Cheng Chi, Shuran Song), ACT (Tony Zhao), DP3 (Yanjie Ze, point clouds) and HPT (Lirui Wang, Kaiming He) have dormant reference repos, but Diffusion Policy and ACT live on in LeRobot, which calls ACT the "recommended first policy".

Closed comparators

• Gemini Robotics 2 (Google DeepMind, Jul 30, 2026): whole-body control and dexterity for partners only. Five-finger tasks succeed 32–92% of the time; On-Device 2 went from 6.7% to 53.3% on an unseen SO-101 arm, and DeepMind says it typically adapts to new bi-arm robots with under 200 examples.
• Generalist AI (GEN-0 to GEN-1.5): trained on 270,000+ hours of real manipulation data growing by 10,000 h a week — two to three orders of magnitude beyond any open dataset.
• Also closed: PI's π*0.6, Figure Helix, 1X Redwood, Skild Brain and Tesla's Optimus models.

Reading the benchmarks

• LIBERO is saturated: at least seven models self-report ≥98%, and LeRobot's π0.5 re-run moved 0.65 points with only 10 trials per task.
• Robustness suites expose memorization: LIBERO-Plus (CVPR 2026) finds models "tend to ignore language instructions completely" and collapse when the camera moves.
• Real-robot evaluation is where models separate: RoboArena (double-blind pairwise trials on DROID setups; π0-FAST-DROID led the paper's seven policies, and GR00T N2 and Cosmos 3 have since both claimed #1) and RoboChallenge (charted below).
• No standard open benchmark exists for humanoid or five-finger manipulation.

Diagram: RoboChallenge Table30 leaderboard snapshot via sota2 (~Apr 2026) · 8 open models, fine-tuned per task (shown in the PDF version).

These are per-task fine-tunes on a 10-robot fleet. Multi-step tasks such as sandwich-making score near 0%, and generalist versions do far worse: the official Oct 2025 report found π0.5 falls from 43.7% task-specific to 17.7% as a single multi-task policy.

Whole-body control, locomotion and motion imitation

Open humanoid control has converged on one pipeline, and its binding constraint is now commercially usable motion data, not algorithms. Nearly every 2025–26 tracker targets the Unitree G1 first, and the strongest one, NVIDIA's SONIC, was trained on 700 hours of private motion capture.

The standard pipeline:

1. Human motion from mocap (AMASS, LAFAN1, BONES-SEED), monocular video (GVHMR, VideoMimic) or live VR and egocentric capture (PICO, TWIST2).
2. Retargeting to the robot's body: GMR (kinematic, fast) or OmniRetarget (keeps contact with objects and terrain).
3. RL motion tracking in GPU simulation — one policy per motion (BeyondMimic) or one generalist (SONIC, TWIST, Humanoid-GPT) — trained in Isaac Lab, mjlab or ProtoMotions.
4. Sim2sim check in MuJoCo.
5. Deployment through ONNX or C++ (often a ROS 2 controller), usually on a G1.

Data quality decides the outcome. In GMR's study, dance clips retargeted by the older PHC method reached 0% tracking success versus 100% with GMR; KAIST's physics-filtered PHUMA data raised G1 tracking success to 92.7% from 76.2% on raw AMASS.

NVIDIA GR00T Whole-Body Control and SONIC

The biggest open release in humanoid control: one universal tracking policy shipped with weights, a planner, teleop and a production C++ stack. Paper

• Who: NVIDIA GEAR; co-first authors Zhengyi Luo (author of PHC), Ye Yuan, Tingwu Wang and Chenran Li; led by Jim Fan and Yuke Zhu.
• Design: A PPO-trained tracker on 700 h of in-house mocap (100M+ frames, 170 subjects), scaled to 42M parameters over ~9k GPU-hours. A motion-token interface accepts PICO VR, three-point VR, video, text, music or VLA outputs; a kinematic planner replans in under 5 ms on a laptop. Training code came out Apr 10, 2026 and a v1.1 checkpoint Jul 23, 2026; code is Apache-2.0 and the weights are commercially usable. The open training recipe uses the public BONES-SEED set (~288 h), not the private corpus.
• Evidence (self-reported): 100% success on 50 varied real-world trajectories; three-point VR teleop with 121.9 ms mean latency and 6 cm mean wrist error; a fine-tuned VLA solved an apple-to-plate task 19 of 20 times.
• Trade-offs: The only open generalist with a deploy stack. It is G1-centric, its training corpus is private, a full retrain costs ~9k GPU-hours, and users report sim2sim falls and G1 over-temperature warnings.

BeyondMimic (UC Berkeley + Stanford)

High-fidelity per-motion tracking that made aerial cartwheels and spin-kicks transfer zero-shot to a real G1. Site · paper

• Who: Qiayuan Liao (lead), Takara Truong, Xiaoyu Huang and Guy Tevet, with Koushil Sreenath and C. Karen Liu.
• Design: An anchor-body tracking reward and joint impedance tuned to 10 Hz, trained in Isaac Lab on LAFAN1 clips and deployed through a ROS 2 Jazzy C++/ONNX controller. The paper's guided-diffusion planning layer was not open-sourced.
• Adoption: Booster adapted it for the K1, HighTorque ported it to the Mini Pi, robot_lab and mjlab reproduce it, and SONIC uses it as a baseline.
• Trade-offs: The best fidelity per clip, but one policy per motion. Issues show Isaac Lab version breakage and painful ports to other robots.

mjlab (UC Berkeley)

Isaac Lab's manager-based API running on MuJoCo Warp, and the fastest-rising lightweight alternative. Paper

• Who: Kevin Zakka, Qiayuan Liao, Brent Yi, Louis Le Lay, Koushil Sreenath and Pieter Abbeel.
• Adoption: 54k PyPI downloads a month. Microduck, Asimov, BFM-Zero (~6 h of training) and Inria's YAHMP study all build on it.
• Trade-offs: One-command install, MuJoCo's contact model, and the same engine for training and sim2sim. It lacks RTX sensors and USD asset tooling, and its FAQ ties some NaN crashes to MuJoCo Warp's beta status.

holosoma and OmniRetarget (Amazon FAR)

The most complete open alternative to NVIDIA's stack across robots and simulators.

• Who: Amazon FAR (Frontier AI & Robotics), including Pieter Abbeel, Rocky Duan, Guanya Shi and Carmelo Sferrazza; OmniRetarget is led by Lujie Yang, Xiaoyu Huang and Zhen Wu.
• Design: Training with PPO or FastSAC, one inference pipeline for sim2sim and sim2real, and contact-preserving retargeting, for the G1 and Booster T1.
• Evidence (self-reported): Sim-to-real locomotion trained in 15 minutes on one RTX 4090; a dance of over two minutes on a real G1; OmniRetarget eliminated foot skating and reached 82% downstream RL success on robot-object interaction.
• Trade-offs: No released generalist checkpoint. Only ~4 h of its ~9 h of retargeted data is public, because retargeted LAFAN1 was withheld "due to licensing restrictions".

GMR (General Motion Retargeting)

The default kinematic retargeter, real-time on CPU. Paper

• Who: Yanjie Ze (now at Amazon FAR), with C. Karen Liu and Jiajun Wu at Stanford.
• Inputs: SMPL-X (AMASS, OMOMO), BVH (LAFAN1), FBX, Xsens, monocular video via GVHMR, and live PICO streaming for TWIST2.
• Trade-offs: Fast and broad, but kinematic only — no dynamics, objects or terrain (use OmniRetarget for those). Users report arm offsets and T-pose calibration issues, and maintainer news stops in Jan 2026.

TWIST and TWIST2 (Stanford + Amazon FAR)

A cheap, fully open loop from humanoid teleop to learned policy. TWIST2's tracker trains on ~20k clips (7k retargeted with GMR); a Diffusion Policy on 224×224 RGB with 2-second action chunks then drives a G1 with Dex3 hands. Code is MIT and the datasets CC BY 4.0. See the teleop section for its capture rig.

ProtoMotions3 and MimicKit

• ProtoMotions3 (NVIDIA Research: Chen Tessler, Xue Bin Peng, Erwin Coumans and others) supports the most backends (Isaac Gym, Isaac Lab 3, Newton 1.0, MuJoCo). It trains on all of AMASS in 12 h on four A100s, and a general tracker trained on BONES-SEED transfers zero-shot to a real G1.
• MimicKit (Xue Bin Peng) offers clean reference implementations of DeepMimic, AMP, ASE, ADD and more, for research and teaching rather than deployment.

Vendor and community stacks

• Unitree unitree_rl_gym (Isaac Gym) and unitree_rl_lab (Isaac Lab 2.3) are the standard start for G1 locomotion; neither covers general motion tracking.
• Booster moved from booster_gym to booster_train, which adapts BeyondMimic for the K1.
• NVIDIA WBC-AGILE is an engineering-grade Isaac Lab 3 workflow for the G1 and T1 with a built-in evaluation framework.
• robot_lab brings ~10 humanoids into Isaac Lab, including OpenLoong and Roboto's ATOM01.
• rsl_rl (ETH RSL) is the PPO library underneath most of these, with 13 PyPI releases in 2026.

Newer generalists and foundation controllers

• Humanoid-GPT (Galbot, CVPR 2026): a GPT-style tracker trained on 2 billion motion frames; only inference code and checkpoints are released.
• BFM-Zero (CMU LeCAR + Meta, ICLR 2026): one policy that can be prompted for tracking, goal-reaching or reward optimization; non-commercial license.
• EgoHumanoid (OpenDriveLab, RSS 2026): loco-manipulation learned from egocentric human demos, reporting +51% over robot-only data (self-reported), with GR00T WBC as its low-level controller.

The 2024–25 generation, now mostly dormant

ASAP (CMU's learned "delta action" sim-to-real correction), HOVER (NVIDIA + CMU), H2O/OmniH2O (non-commercial), PHC (the AMASS-scale baseline), GMT (checkpoint only), HumanPlus, Humanoid-Gym and legged_gym. Many flagship repos, including TWIST2, HOMIE, FALCON and PBHC, still require Isaac Gym on Python 3.8.

Model-based libraries

These are healthy but now serve mostly as IK, retargeting and hybrid-control components rather than complete humanoid stacks.

• Pinocchio (Inria and LAAS-CNRS) is the rigid-body dynamics base layer; Crocoddyl and Aligator add trajectory optimization; TSID, Pink and PlaCo add task-space IK and inverse dynamics; mink does differential IK on MuJoCo.
• Drake (TRI), MuJoCo MPC (Google DeepMind), OCS2 (ETH) and IHMC Open Robotics Software (Java whole-body control from the Valkyrie and Nadia lineage) cover MPC and verification.
• Integrated open MPC+WBC humanoid stacks (legged_control, OpenLoong-Dyn-Control) are discontinued or dormant. Hybrids are returning: HybridMimic (Booster T1, 2026) lets an RL policy modulate a centroidal controller and reports 13% lower tracking error than pure RL.

Simulators, physics engines and training frameworks

Rigid-body simulation consolidated in 2025–26. Newton (a Linux Foundation project started by NVIDIA, Google DeepMind and Disney) and MuJoCo Warp became the shared GPU core, Isaac Lab 3.0 can run without Isaac Sim, and Meta and NVIDIA retired Habitat and Isaac Gym. The defensible value has moved up the stack: assets, scenes, calibration and evaluation.

NVIDIA's own Isaac Lab benchmark (Sep 2026; 8,192 environments on an RTX PRO 6000 Blackwell; end-to-end training FPS) shows the Newton backend roughly 1.5–2× faster than PhysX:

NVIDIA ran these itself; the rows are flagged "dirty working tree", Newton runs often vary >25% between runs, and contact parameters differ across backends.

MuJoCo, MuJoCo Warp, Menagerie and Playground (Google DeepMind)

The reference engine for contact-rich robotics, now with a GPU twin.

• MuJoCo: 4.0M PyPI downloads a month and monthly releases since 3.5 (Feb 2026), which added a system-identification toolbox and actuator and sensor delays. 3.14 (Sep 22, 2026) added experimental penetration-free contact for deformables.
• MuJoCo Warp: "officially released" in Feb 2026 as an NVIDIA-GPU rewrite with a batch renderer that draws Gaussian splats. It is the main solver behind Newton, Isaac Lab 3's Newton mode, mjlab and Playground (383k downloads a month). Its docs warn it scales poorly beyond ~60 DoF in one kinematic tree — a G1 with two dexterous hands exceeds that.
• Menagerie: 71 curated robot models, including 11 humanoids and 6 dexterous hands, now pip-installable; licenses vary by model.
• Playground (RSS 2025): GPU environments that "train policies in minutes on a single GPU", with zero-shot sim-to-real on humanoids, quadrupeds and hands.
• Trade-offs: Best accuracy, determinism and fewest dependencies. Basic rendering, and USD only through converters.

Newton (Linux Foundation)

The new shared GPU physics engine, with one API over interchangeable solvers.

• Who: Started by NVIDIA, Google DeepMind and Disney Research. Its steering committee includes Yuval Tassa (DeepMind), Miles Macklin (NVIDIA), Disney's Vassilios Tsounis and Moritz Baecher, and TRI's Drake developers.
• Releases: 1.0 on Mar 10, 2026, then monthly releases to 1.6 on Sep 10, 2026.
• Design: MuJoCo Warp and Featherstone for articulated robots; XPBD, VBD (cloth and soft bodies) and MPM (granular); experimental hydroelastic contact; USD, MJCF and URDF import.
• Trade-offs: The broadest open multi-solver engine with neutral governance. APIs change every minor release, several solvers are experimental, and speed requires NVIDIA GPUs.

NVIDIA Isaac Sim, Isaac Lab and Isaac Lab-Arena

The default stack for humanoid sim-to-real, now modular.

• Isaac Sim: source code Apache-2.0 on GitHub since 5.0 (Aug 2025), but building it needs NVIDIA's proprietary Omniverse Kit SDK. 6.0 (Jun 2026) added NuRec Gaussian-splat scenes; 6.1 followed in Sep 2026. Minimum GPU: RTX 4080 or A40.
• Isaac Lab: BSD-3, 16+ robots and 30+ environments. Version 3.0 (Early Access Sep 16, 2026) splits physics (PhysX, Newton, OVPhysX) and rendering into plug-ins and can run without Isaac Sim. Its Mimic data-generation module requires NVIDIA's proprietary cuRobo.
• Isaac Lab-Arena (alpha): builds environments from a scene, an embodiment and a task, with experimental natural-language environment generation and perturbation sweeps — the same space an environments startup would target.
• Adoption: Isaac Lab has 3.9k forks; unitree_rl_lab, BeyondMimic, holosoma and RoboTwin build on it.
• Trade-offs: The richest sensors and assets. Heavy installs, a high GPU floor, breaking 2.x → 3.0 changes and mixed licensing.

Genesis World (Genesis AI)

The one VC-backed independent simulator, and the only fast option across GPU vendors.

• Who: Began as an academic project in Dec 2024; now run by Genesis AI (CEO Zhou Xian), which raised a $105M seed in Jul 2025 led by Eclipse and Khosla Ventures. Version 1.0 shipped May 27, 2026, with roughly weekly releases since.
• Design: Rigid, FEM, MPM, SPH, PBD and IPC solvers in one scene; NVIDIA, AMD, Apple and Vulkan GPUs through its own compiler (Quadrants); tactile sensors.
• Controversy: Its Dec 2024 README claimed 43M FPS (430,000× real time) for a Franka arm on one RTX 4090. ManiSkill's Stone Tao argued the number came from a near-trivial setup; the claim was dropped, and CI benchmarks now gate contributions. No independent head-to-head of Genesis 1.x was found.
• Strategy: The company has gone "full stack" with its own hand (with Wuji), a data glove and the GENE-26.5 model; its CEO says simulation's main value is evaluation.
• Trade-offs: The widest multiphysics and cross-vendor support, but damaged credibility on speed and a smaller ecosystem.

ManiSkill3 / SAPIEN (UCSD + Hillbot)

Hao Su's GPU-parallel manipulation simulator, notable for running a different scene in every parallel environment. It claims 30,000+ FPS with rendering and, in an older self-run comparison, 2.2× Isaac Lab's throughput on visual CartPole at a quarter of the memory. Easiest pip install and efficient visual RL; Linux-only GPU mode and thin humanoid coverage.

Drake, Gazebo, Webots and Chrono

• Drake (TRI; monthly releases) is best for verification and model-based manipulation with hydroelastic contact, not massively parallel RL.
• Gazebo (Jetty LTS, supported to 2031) is the standard for ROS integration testing, not GPU training.
• Webots (Cyberbotics) last tagged a release in early 2025 and suits teaching.
• Project Chrono (UW–Madison; 10.0 in Mar 2026) covers deformable terrain, vehicles and granular media.
• PyBullet is legacy, and Brax now maintains only its training library.

Benchmark environments and scene suites

• RoboCasa365 (UT Austin, Yuke Zhu; v1.0 Feb 2026): 365 tasks, 2,500+ kitchen scenes, 3,200+ objects, 600+ h of human demos and 1,600+ h of generated robot data; assets CC BY 4.0, CPU MuJoCo.
• BEHAVIOR-1K / OmniGibson (Stanford; Fei-Fei Li, Jiajun Wu): 1,000 household activities on Isaac Sim with fluids, cloth and object states.
• RoboTwin 2.0 (ICML 2026): a bimanual benchmark and data generator; its paper reports domain-randomized synthetic data plus 10 real demos improved success 367% (relative).
• RoboVerse puts one API over 8+ simulators; Gymnasium-Robotics, Meta-World, MyoSuite and dm_control remain maintained.

Procedural scenes, real-to-sim twins and world models

• Infinigen (Princeton; 2.0 alpha Jul 2026) generates procedural scenes and sim-ready articulated assets in Blender — the most credible open generator, though outputs need physics tuning per simulator. The 2023–24 LLM generators (Holodeck, RoboGen, GenSim) are dormant.
• Gaussian-splat twins: Isaac Sim's NuRec, MuJoCo Warp, Newton and Genesis all render splats. GaussGym (Berkeley) claims 100,000+ steps/s on consumer GPUs with iPhone-scanned and Veo-generated scenes; DISCOVERSE and GS-Playground pair splats with physics. Splats carry appearance only — not mass, friction or articulation.
• World models as simulators: NVIDIA Cosmos 3 is open and commercially usable; 1X's open challenge offers 100+ h of EVE robot video; Google's Veo-based policy evaluator (validated on 1,600+ real trials) and Genie 3 are closed.

What it means: rigid-body physics is becoming a free, well-funded commodity, so building another solver is a poor bet. The gaps are sim-ready assets in both USD and MJCF with tuned physics, procedural scene and task generation, calibrated actuator models, real-to-sim twins that include physics, and evidence that simulated results predict real ones.

Middleware and tooling

LeRobot has become the de facto robot-learning toolchain and dataset format — 77,000+ Hub datasets are tagged LeRobot — while ROS 2 remains the systems backbone. 2026 brought dora-rs 1.0, Copper 1.0, ROS 2 Lyrical with zero-copy GPU buffers, and Qualcomm's deal to buy PickNik, MoveIt's steward.

LeRobot (Hugging Face)

The end-to-end PyTorch library, dataset format and Hub ecosystem that most open robot learning now runs through. GitHub · paper

• Who: Hugging Face's robotics team in Paris. Original lead Rémi Cadène left to found UMA (Paris), which is building the "Northstar" humanoid with LeRobot co-founder Simon Alibert, SO-100 designer Robert Knight and ex-DeepMind researcher Pierre Sermanet; Thomas Wolf advises it. Releases are now cut by Steven Palma with Caroline Pascal, Pepijn Kooijmans and others.
• Hardware: Native support for SO-100/101, LeKiwi, Koch, HopeJR, OMX, Reachy 2, OpenArm, reBot and the Unitree G1 (whole-body control since v0.5). Plugins are auto-discovered for xArm, UR5e, Franka, Piper, WidowX, ARX5, I2RT YAM, GELLO, Quest, ROS 2 bridges and tactile cameras.
• Policies: ACT, Diffusion, VQ-BeT and Multitask DiT; HIL-SERL and TD-MPC for RL; π0, π0-FAST, π0.5, GR00T N1.7, SmolVLA, X-VLA, EO-1, MolmoAct2 and WALL-OSS; world models (VLA-JEPA, LingBot-VA, FastWAM) and reward models (SARM, TOPReward, Robometer).
• Dataset format v3.0 (Oct 2025): many episodes packed per Parquet/MP4 file, streaming straight from the Hub, and an optional LanceDB backend. v0.6 (Jul 2026) added depth, video trim and re-encode tools, language columns with a VLM subtask-annotation pipeline, recording of evaluation rollouts, DAgger takeovers, cloud training on HF Jobs and Foxglove visualization. Benchmarks plug in through "EnvHub" (LIBERO, MetaWorld, RoboCasa365, RoboTwin 2.0, VLABench and more).
• Adoption: 238k PyPI downloads a month; an ICLR 2026 paper; its Jun 2025 worldwide hackathon drew 3,000+ participants in 44 countries (self-reported).
• Trade-offs: Learning-centric and Python-first, not a real-time middleware. Every release breaks something (v0.5 requires Python 3.12; v0.6 renamed modules), and v2.1 → v3.0 conversions have corrupted episode boundaries in the wild.

ROS 2 and its ecosystem

• ROS 2 Lyrical Luth (May 22, 2026; LTS to May 2031) targets Ubuntu 26.04 and adds zero-copy rosidl::Buffer publishing with pluggable GPU backends. rmw_zenoh has been Tier 1 since Kilted, and MCAP has been the default bag format since Iron.
• MoveIt 2 (BSD-3): Qualcomm agreed on Sep 23, 2026 to acquire PickNik and pledged to keep MoveIt open source; MoveIt Pro stays proprietary.
• ros2_control is the standard hardware abstraction; BeyondMimic's real-robot controller is built on it.
• NVIDIA Isaac ROS 5.0 (Sep 21, 2026) moved to Lyrical and replaced its NITROS transport with native GPU buffers; 4.6 added cloud-control workflows for the Unitree G1.

Lightweight runtimes

• dora-rs 1.0 (Sep 2, 2026): YAML-defined dataflow graphs of Rust, Python and C++ nodes exchanging Apache Arrow messages, zero-copy over Zenoh shared memory. It claims 10–17× lower latency than ROS 2's rclpy (self-reported) and is used by LeRobot Humanoid and OpenArm's data collection.
• Copper-rs 1.0 (Jul 2, 2026): a deterministic, compile-time-scheduled Rust runtime with structured logging and exact replay.
• Zenoh (Eclipse Foundation, backed by ZettaScale): pub/sub that survives Wi-Fi and WAN links; it underpins rmw_zenoh and dora.

Visualization and logging

• Rerun (Stockholm; ~$17M seed in Mar 2025): multimodal logging and a viewer; LeRobot's dataset visualizer is built on it.
• MCAP (MIT): the self-describing log container and ROS 2 default. Foxglove's app went proprietary in 2024; BMW-led Lichtblick (MPL-2.0) continues the open Foxglove Studio code.
• Viser (Brent Yi and collaborators): browser-based 3D GUIs from Python, with 837k downloads a month.

Assets and data plumbing

• robot_descriptions.py (Stéphane Caron): one-line imports of 100+ URDF and MJCF models; 3.2.0 shipped Sep 12, 2026.
• any4lerobot: converts Open X-Embodiment, AgiBot World, RoboMIND, LIBERO and RoboCasa to LeRobot and migrates between v2.1 and v3.0.
• LanceDB and FiftyOne: general data tools now used for robot datasets (LeRobot's Lance backend; Voxel51 hosts Egocentric-10K subsets).
• trajlens: a new LeRobot dataset linter; its author's audit of 100 public LeRobot datasets found 81% had issues or failed to load cleanly.
• Fading: phosphobot (no release since Oct 2025), RLDS (the legacy Open X-Embodiment format) and YARP (iCub only).

Datasets and benchmarks

Open robot data is growing fast but remains small, noisy and mostly non-commercial. The largest real-robot and hand datasets forbid commercial use, 2026's big open releases are egocentric human video, and evaluation is shifting from saturated simulation suites to real-robot and long-horizon tests.

Robot datasets

Egocentric human data

EgoVerse found human data improves robot policies only when it is aligned with the robot's objective; raw volume alone does not help.

Human motion data for humanoid control

Benchmarks and competitions

The data market

Money is flowing into robot-data suppliers, and prices are already compressing.

• Suppliers (funding as reported): XDOF ($70M; co-founded by GELLO's Philipp Wu and Fred Shentu with Nemo Jin; about 20 customers including frontier labs), Scale AI (Physical Intelligence and Generalist as customers), Lightwheel (SimReady assets plus EgoSuite; customers include Google DeepMind and Figure), Mecka AI (about $68M; body-worn sensors; 1X as a customer), Config ($27M seed at a $200M valuation), Human Archive ($8.2M; YC W26), Build AI (about $15M), Encord ($60M Series C; curation tooling), Bones Studio (studio mocap), Luel (rights-cleared marketplace), Axis Robotics ($12M; crowdsourced sim teleop), Genesis AI (a closed tactile data glove) and MANUS (gloves).
• Published prices (vendors, Jul–Sep 2026): $15–40 per raw hour of egocentric video; $28–60 per hour of teleop; $8–25 per hour extra for annotation; $80–150 per hour for full humanoid multi-sensor capture. Raw footage resells for $2–5 per hour, buyers discard ~90% of it, and collectors in India earn $1–4.20 per hour.
• Formats: LeRobotDataset (v3.0) for training and distribution; MCAP for raw logs; legacy RLDS, HDF5, WebDataset and Aria VRS elsewhere.

Trade-offs and recommended stacks

If you are starting today, the safe defaults are a Unitree G1 or SO-101/OpenArm for hardware, Isaac Lab or mjlab on MuJoCo or Newton for simulation, LeRobot v3 for data, and π0.5, GR00T N1.7 or SmolVLA as the base model. The real decisions are licensing and how much of the hardware you want to own.

Where the open ecosystem is thin

These are the gaps a data, simulation or environment company could fill, with the evidence behind each.

1. Commercially licensed data with provenance. The largest open robot, hand and motion datasets (AgiBot World, Galaxea, ActionNet, EgoDex, AMASS, LAFAN1) are non-commercial. NVIDIA said GR00T's earlier non-commercial weights were "a data issue, not a license choice".
2. Hand data with force and touch. Open gloves are research-grade or license-restricted; commercial gloves cost $1.5k–$7.3k; open dexterous datasets hold ~40–100 demos per task; every mainstream VLA benchmark uses parallel grippers.
3. Egocentric video with action labels. Video-only egocentric data is already a commodity (Egocentric-100K, raw footage resold for a few dollars an hour). What is scarce is measured hand and wrist pose, depth and coverage aligned to target skills — the ingredients behind GR00T N1.7's 20,000-hour EgoScale pretraining and RDT2's 10,000 hours of UMI data.
4. Humanoid motion and whole-body data. The best open tracker (SONIC) was trained on 700 h of private mocap, the largest robot-ready set (BONES-SEED) is gated, and curation visibly pays: PHUMA's physics filtering and GMR's retargeting each move tracking success by double digits.
5. Sim-ready assets and calibrated actuator models. Menagerie misses most open humanoids and hands, every cheap-robot project builds its own system identification, and Gaussian-splat twins carry appearance but no physics.
6. Neutral evaluation. LIBERO is saturated, RoboChallenge is co-run by a model maker, RoboArena needs thousands of episodes, and no open benchmark exists for humanoid or five-finger manipulation. OpenArm built its Cell because cross-lab results do not compare.
7. Data quality and format plumbing. An audit found 81% of 100 public LeRobot datasets had issues; v2.1 → v3.0 conversions corrupt episodes; teams juggle MCAP, LeRobot, HDF5 and RLDS.

Who is already there: NVIDIA (Isaac Teleop for robot-free egocentric capture, Isaac Lab-Arena for environment generation, GRAIL data, Cosmos), Hugging Face (LeRobot's annotation pipeline, RoboChallenge), Lightwheel (sim assets plus EgoSuite), XDOF, Scale AI, Mecka AI, Build AI, Human Archive, Genesis AI (tactile glove), Bones Studio (mocap), Encord (curation), Luel (rights-cleared data) and Axis Robotics (crowdsourced sim teleop). Differentiation will have to come from sensors, quality assurance, rights and robot-specific alignment rather than raw volume or teleop software.

Stalled, archived or discontinued

Projects stall for three reasons: companies run out of money (K-Scale Labs), labs move on after the paper (most 2024 robot-learning repos), and platform owners retire tools (Isaac Gym, Habitat, ODrive's open firmware). Many are still worth reading or reproducing; few are worth building on.

Methodology and sources

• What counts as open: design files, code or model weights published under a license that allows reuse; partially open projects (open software on closed hardware, non-commercial weights) are flagged. Closed products appear only as comparators.
• Status rules (as of Oct 2, 2026): Active = commits, releases, new hardware revisions, papers or shipping since about April 2026; Slowing = last dated activity 6–12 months ago; Dormant = over 12 months; Discontinued = officially ended or the company closed.
• How it was gathered: parallel research passes on Oct 2, 2026 over primary sources — GitHub repo and release pages, PyPI, Hugging Face model and dataset cards, arXiv papers, vendor pages and reputable press. The GitHub API was unavailable, so star counts come from GitHub pages or a package index and are rounded, and status leans on release, PyPI and paper dates where commit dates were not visible.
• Caveats: vendor and author numbers are marked "self-reported"; benchmark results are mostly self-reported; Hugging Face download counts are rolling file requests that differ up to 30× between views; several Chinese projects develop on AtomGit or Gitee mirrors that were not fully checked; hardware unit sales are rarely published. Some pages were rate-limited, and facts seen only in search headlines are marked "reported".

Sources

Hardware and hands

• Berkeley Humanoid Lite · ToddlerBot paper · Menlo Research: humanoid legs in 100 days · Roboto Origin coverage · K-Scale Labs shutdown · HopeJR announcement · Unitree G1
• LEAP Hand paper · ORCA v1 · ORCA platform · RUKA · RUKA-v2 · Aero Hand Open · RAPID Hand · Sharpa Wave listing
• UMI · FastUMI-100K · DexUMI · DOGlove · Taccel · Sparsh

Arms and teleop

• SO-101 launch (TechCrunch) · OpenArm · OpenArm 2.0 listing · reBot B601 · GELLO paper · AgileX Piper · Trossen WidowX

Models and benchmarks

• openpi · Physical Intelligence (Wikipedia) · GR00T N1.7 card · NVIDIA forum on GR00T licensing · NVIDIA GTC 2026 robotics news
• OpenVLA-OFT · SmolVLA blog · Xiaomi-Robotics-0 · MolmoAct2 · AI2 MolmoAct2 blog · Gemini Robotics 2 coverage · Generalist GEN-0
• LIBERO-Plus · LIBERO-PRO · RoboChallenge report · RoboChallenge snapshot (sota2) · RoboArena · BEHAVIOR 2025 winner · BEHAVIOR Challenge · RoboTwin 2.0 · ArmnetBench · AgiBot World Challenge 2026 · World Humanoid Robot Games 2026 · RoboCup 2026 results

Whole-body control and motion data

• SONIC · BeyondMimic · GMR: Retargeting Matters · holosoma paper · OmniRetarget · TWIST2 · PHUMA · AMASS license · BONES-SEED

Simulation

• Isaac Lab report · Isaac Lab releases · Newton · MuJoCo · MuJoCo Playground · mjlab · ManiSkill3 · GaussGym

Tooling, datasets and the data market

• LeRobot paper · LeRobot dataset analysis · LeRobot dataset lint audit · UMA and Rémi Cadène · Qualcomm acquires PickNik · EgoSuite-Open100K
• XDOF (TechCrunch) · Human Archive (TechCrunch) · DreamVu: robot data companies 2026 · EXYLOS dataset pricing · DataX Power collection costs · Flikforge on the egocentric-video glut