Files: MD

Contents

Working notes for Open-Source Humanoid Robotics Landscape (Oct 2026), compiled Oct 2, 2026. Not fact-checked line by line: where these notes and the report disagree, trust the report. See README.md.

Open-source robot arms, bimanual/mobile rigs and teleoperation for robot-learning data (status as of 2026-10-02)

How this was researched, and the access limits

  • All research was done on 2026-10-02. Partway through, the shared session ran out of WebSearch quota (200 of 200 calls). WebFetch was also repeatedly blocked by the shared proxy (HTTP 429 rate limits, then the 800-per-hour budget ran out). To work around this, most repository facts were read from the GitHits code index, which holds default-branch snapshots from about October 2026. These are marked [idx].
  • Star counts are marked by source:
    • [gh]: read from the GitHub page with WebFetch today.
    • [pkg]: repository metadata attached to the PyPI package via GitHits, refreshed 2026-10-02 and rounded.
  • Pages I could not read because of rate limits or budget:
    • CNX Software "OpenArm 2.0" (2026-09-30)
    • github.com/enactic/OpenArm and /releases
    • Seeed's reBot B601 launch blog (2026-04-20)
    • docs.openarm.dev (the root page and hardware/openarm-2.0/general)
    • humanoid.guide OpenArm 02
    • kamenski.me's LeRobot dataset analysis
    • github.com/TheRobotStudio/SO-ARM100 (stars taken from the org page instead)
    • huggingface.co/datasets?other=LeRobot
    • the GELLO project site
    • github.com/LightwheelAI/leisaac
    • huggingface.co/datasets?search=openarm
  • Searches I could not run because of the exhausted budget: SO-101 unit sales, hackathon participant counts, AR4-MK4 details, and Franka/Kinova/UR list prices.
  • How the Hugging Face dataset counts work. "tag" means huggingface.co/datasets?other=X; "name" means ?search=X, a substring match on the repo name that can include unrelated datasets. Tags are added by users and are inconsistent; that inconsistency is itself a finding (see the gaps section). All counts were fetched 2026-10-02.

Headline takeaways for a data, simulation or environment startup

  1. Two clusters have formed.
    • The low-cost cluster is SO-101 and its derivatives (LeKiwi, XLeRobot, AlohaMini, AM-ARM). It dominates dataset volume: 9,432 Hugging Face datasets have "so101" in the name, and SO-100s make up 31.4% of LeRobot's curated community dataset.
    • The quasi-direct-drive (QDD) / CAN-bus mid-tier, roughly $2k–$7k, is where research-grade data is moving. It includes OpenArm 2.0, Seeed reBot B601, I2RT YAM, AgileX Piper and Trossen's 2026 WidowX. Each of these has very few public datasets (for example, only 39 tagged openarm).
  2. The open data commons is tiny and noisy. LeRobot's curated community_dataset_v3 is 791 datasets, 235 contributors and 251.5 hours in total. HF's own blog lists empty task strings, ambiguous camera keys and inconsistent action/state dimensions.
  3. Teleop software is commoditising quickly, much of it driven by large players. Examples: NVIDIA Isaac Teleop (Apache-2.0, v1.4.145 on 2026-09-17, positioned as "the unified framework for high-fidelity ego-centric and robotics data collection"), LeRobot v0.6 (DAgger rollout, phone and VR teleop, the LeLab GUI), Trossen (Quest VR teleop, a UMI-style gripper, cloud tooling), and Enactic (the KER leader, its own dataset format).
  4. Calibration and sim fidelity of low-cost arms is a measurable gap.
    • The SO-101 maintainers' own document says hand-held calibration introduces ±50–200 ticks of error per joint (out of 4096 ticks per revolution).
    • There are two calibration conventions in the URDF/MJCF files.
    • The LeRobot gripper mapping is "not yet reflected" in the sim files.
    • The OpenArm Cell exists specifically to remove assembly tolerance and per-lab drift from data and evaluation.
  5. Licensing traps.
    • Non-commercial: HOMIE and AirExo-2 (CC BY-NC-SA).
    • No license: Dummy.
    • Copyleft: PAROL6 (GPL-3.0); OpenArm hardware (CERN-OHL-S, strongly reciprocal).
    • Gated: GELLO CAD is available on request only.
    • Seeed moved reBot from a non-commercial license to CERN-OHL-W in May 2026, which signals a trend toward commercial-friendly terms.

A. Open-source arms

SO-100 / SO-101 (TheRobotStudio × Hugging Face) — the default $100–$500 arm for LeRobot

  • Status: ACTIVE.
    • LeRobot v0.6.0 (2026-07-06) and v0.6.1 (2026-08-03) ship unified so_follower/so_leader types; SO-100 and SO-101 were consolidated in v0.5.0.
    • The v0.6.0 release added an NVIDIA Isaac Teleop → SO-101 example, and the LeLab GUI "currently supports the SO-ARM101".
    • Hardware: no new revision since SO-101 (April 2025). The repo CHANGELOG's last entry is v0.1.13 (2025-01-27). The README has since gained more vendors, a WebUI 3-point calibration tool, and SO-101 URDF/MJCF files [idx].
    • I found no official successor ("SO-102") in the repo or in LeRobot's code.
  • Who. The Robot Studio (France) designs the hardware. The Hugging Face LeRobot team maintains the software; v0.6.0 blog authors include Steven Palma, Pepijn Kooijmans, Caroline Pascal and Thomas Wolf. HF calls SO-101 "our signature robot arm". Clem Delangue (28 April 2025): "SO-101 are the first robot arms any AI builder should buy."
  • Links. github.com/TheRobotStudio/SO-ARM100 · github.com/huggingface/lerobot · huggingface.co/docs/lerobot/so101
  • GitHub stats.
    • SO-ARM100: 7.2k stars, 650 forks [gh, org page].
    • huggingface/lerobot: 27k stars, 5.7k forks, 976 open issues [pkg].
    • lerobot on PyPI: 238k downloads/month; latest release 0.6.1 on 2026-08-03 [pkg].
  • License. Apache-2.0 for both hardware/docs (SO-ARM100) and software (LeRobot).
  • What it is.
    • A 5-DoF arm plus a 1-DoF gripper, built from six Feetech STS3215 TTL bus servos through a Waveshare bus-servo board, connected to the host over USB-C.
    • Follower: six STS3215 7.4 V servos with 1/345 gearing, rated 16.5 kg·cm stall at 6 V. An optional 12 V version gives 30 kg·cm and needs a 12 V/5 A supply.
    • Leader: mixed gear ratios (one 1/345, two 1/191, three 1/147) so it can be backdriven, with a trigger handle.
    • Parts are PLA+ prints at 15% infill.
    • Official reach, payload and repeatability are not published in the repo.
  • Price and build.
    • Self-sourced bill of materials (US, excluding printing): $121.94 for one follower; $229.88 for a leader+follower pair [idx].
    • Kits sell for $100–$500 depending on vendor, assembly and tariffs (TechCrunch, 2025-04-28).
    • Seeed SO-ARM101 Pro servo kit: $277.99 (no printed parts).
    • WowRobo lists $199 (all variants sold out on 2026-10-02).
    • Vendors listed in the README: RobotEd (CH), Robonine, PartaBot (US), ForgeMotion (US), Seeed (international, China, Japan), WowRobo, RoboSEasy (KR), NeoBot (CN), Autodiscovery (EU). phospho sells an SO-100 follower.
    • Build time is not stated officially.
  • ROS / LeRobot / simulation.
    • LeRobot is the primary stack. ROS 2 access is through third-party plugins (leros2, ROBOTIS lerobot_robot_ros2_zenoh).
    • The repo has URDF for SO-100 and URDF plus MJCF for SO-101. MuJoCo Menagerie has trs_so_arm100 and robotstudio_so101. Isaac Lab is supported through LeIsaac.
  • Adoption.
    • No official unit-sales figure has been published.
    • Hugging Face datasets: 9,432 with "so101" in the name, 4,033 with "so100" in the name; 4,890 tagged so100, 1,237 tagged so101, 15 tagged so101_follower.
    • LeRobot community_dataset_v3: SO-100 series = 31.4% of 791 datasets.
    • phospho says "2,000+ robots using phosphobot" (self-reported; mostly SO arms).
    • Events: SEEED×NVIDIA×HF "Embodied AI Home Robot Hackathon" (25–26 Oct 2025, Bay Area, per the XLeRobot README). A "LeRobot Worldwide Hackathon – Munich, Jun 13 (2026)" listing appeared in search results (not opened).
  • Benchmarks.
    • SmolVLA on real SO-100 tasks: 51.7% success without community-data pretraining, 78.3% with it; about 78% in both synchronous and asynchronous inference (HF SmolVLA blog).
    • LeRobot compute guide: ACT on about 50 episodes (45k frames) trains in roughly 30–60 minutes on an RTX 4090.
  • Trade-offs.
    • Strengths: the cheapest entry into the largest dataset and model ecosystem (LeRobot, LeLab, LeIsaac, Isaac Teleop, SmolVLA), Apache licensing, and many vendors.
    • Weaknesses:
      • Hobby servos mean low payload, backlash and no torque sensing.
      • Calibration noise: the maintainers' WebUI-calibration document says hand-held midpoint calibration gives "±50–200 ticks per joint" and can drive the elbow into its hard stop. It caps torque at 30% "to prevent gear stripping".
      • Simulation: the URDF/MJCF ship two calibration conventions (old and new), the LeRobot gripper 0–100 mapping is "not yet reflected", motor parameters are borrowed from Open Duck Mini, and base collision meshes were removed.
    • Choose SO-101 for cheap data volume, teaching and VLA prototyping. Choose OpenArm, reBot, Piper or YAM for payload, stiffness or contact-rich data.
  • Sources.

OpenArm (Enactic) — a human-scale 7-DoF QDD arm; 2026 brought v2.0 plus the KER leader and the Cell

  • Status: ACTIVE.
    • Enactic's org repos were updated 24–29 September 2026 [gh].
    • openarm_dataset released 0.2.0 (2026-06-12), 0.3.0 (2026-06-30) and 0.4.0 (2026-07-28) [pkg].
    • KER was teased on 2026-03-03 and announced as released on 2026-06-26 ("marks the completion of the OpenArm 2.0 ecosystem"). These dates come from X post IDs seen in search results.
    • OpenArm 2.0 was announced around March 2026 (a robohorizon URL is dated 2026/03) and covered by CNX on 2026-09-30 (search listing only).
  • Who. Enactic, Inc. (Tokyo). Co-founder and CEO Hiroto Yamamoto, per a LinkedIn title seen in search results. RT Corporation is the official Japan partner; WowRobo (Shenzhen) is the certified manufacturer. I found no funding information.
  • Links. github.com/enactic/openarm (documentation-site source) · enactic/openarm_hardware · openarm_can · openarm_description · openarm_mujoco · openarm_dataset · openarm.dev
  • GitHub stats.
    • enactic/openarm: 3,534 stars, 378 forks, updated 2026-09-29 [gh].
    • openarm_mujoco 61; openarm_mjlab 12; openarm_dataset 11 stars / 12 forks; openarm_control 5; dora-openarm-data-collection 4.
    • The org has 51 repos and 237 followers.
  • License. Hardware: CERN-OHL-S-2.0 (strongly reciprocal), from openarm_hardware. Software: Apache-2.0 (openarm, openarm_description, openarm_dataset, openarm_can).
  • What it is.
    • v1.0/1.1.
      • Seven DoF per arm, proportioned for a 160–165 cm person.
      • Backdrivable QDD Damiao motors: J1–J2 DM8009, J3–J4 DM4340, J5–J7 and the gripper DM4310.
      • CAN-FD at 1 Mbps nominal / 5 Mbps data. Linux only, because the CAN adapter has no macOS driver.
      • Payload 6.0 kg peak / 4.1 kg nominal per arm. Aluminium and stainless-steel structure.
      • "$6,500 USD for a complete bimanual system" (LeRobot docs).
      • Teleop uses a powered leader arm with bilateral or unilateral control.
    • v2.0 (2026).
      • Same motor lineup (DM-J4310-2EC, DM4340, DM-J8009P) and payload envelope.
      • New compact gripper with an in-hand camera and replaceable fingers.
      • WowRobo listing: 606 mm reach per arm, 5.5 kg per arm, "1 kHz CAN-FD real-time control", optional ZED stereo head camera.
    • KER (Kinematic Equivalent Replica).
      • A motorless leader arm worn like a backpack: magnetic encoders and bearings only.
      • Same joint structure as v2.0 with link lengths scaled to 70%, giving a 1:1 joint mapping with no retargeting.
      • Weighs 1.7 kg (carbon-fibre tube); can be put on in 30 seconds. Designed to reduce operator fatigue in sessions lasting hours.
    • OpenArm Cell. A standardized evaluation enclosure: MISUMI frame, fixed lighting and cameras, adjustable Z axis, an area-sensor stop, and a zero-position calibration jig. In Enactic's words, it exists so that "Model A outperforms Model B" comparisons carry meaning across labs.
    • OpenArm Mini. A seven-DoF-plus-gripper leader built from STS3215 servos, available in LeRobot as openarm_mini / bi_openarm_mini.
    • Software.
      • openarm_can: C++ library and CLI, with packages for Ubuntu 22.04/24.04/26.04 and the RHEL family. Python bindings are marked "UNSTABLE".
      • ROS 2, MuJoCo (openarm_mujoco), mjlab, an Isaac Lab page, and LeRobot core support (openarm_follower/openarm_leader, plus bimanual).
      • Data collection and WebXR teleop based on dora-rs.
      • openarm_dataset, Enactic's own format, with validate/repair/merge tools and converters to LeRobot v2.1/v3.0 and GR00T.
  • Prices. From the vendor table in the docs source (October 2026), apparently for bimanual sets:
    • WowRobo: $5,400 (V1.1), $6,500 (V2), $6,200 (Cell), $2,599 (KER), $49 (KER encoder unit). WowRobo's shop shows V2 at $6,000–$6,500 without the stereo camera and $6,800–$7,300 with it; lead time 20–30 days.
    • Anvil Robotics (Taipei): $5,600 for V2, "ships within 48 hours"; notes "OpenArm 1.0 deprecated".
    • Cereboto: $6,280 / $7,080 (V2 without/with camera), $5,000 (V1), $1,000 (V1→V2 upgrade), $2,399 (KER).
    • VLAI $4,699; SVTRobotics $5,800; Soma $6,430 (CNY 45,800); others $4,999–$5,399; RT Corporation by quotation.
  • Adoption.
    • 3.5k stars; at least 10 listed manufacturers; first-class LeRobot support.
    • Only 39 Hugging Face datasets are tagged openarm.
    • HF's own LeRobot org hosts "openarms-hardware-modifications".
  • Benchmarks. None independent found. The VR teleop prototype runs "~20 FPS with TCP" and, per the docs, only in Isaac Lab simulation so far.
  • Trade-offs.
    • Strengths: the only fully open (CERN-OHL-S plus Apache) human-scale 7-DoF arm with real payload and backdrivability. That enables bilateral teleop and force-rich data. It also comes with a full data stack (KER, Cell, dataset tools).
    • Weaknesses: Linux-only CAN; dependence on Damiao motors; a bespoke dataset format; hardware reciprocity under CERN-OHL-S; a small dataset and model base.
    • Compared with AgileX Piper ($1,999, closed, 6-DoF), choose OpenArm for 7-DoF kinematics or open hardware. Compared with Franka FR3, it lacks joint torque sensing and industrial reliability, but costs a fraction.
  • Sources.

Seeed reBot Arm B601 (DM and RS variants) plus the Star Arm 102 leader — the open "step-up" desktop arm of 2026

  • Status: ACTIVE.

    • Launch posts are dated 2026-04-17 (CNX) and 2026-04-20 (Seeed blog) according to their URLs; I did not open them.
    • The README's community showcase links posts dated 2026-04-21 to 2026-09-28.
    • It has LeRobot core support: rebot_b601_follower, bi_rebot_b601_follower, rebot_102_leader.
  • Who. Seeed Studio (Seeed-Projects). The leader arm is FashionStar's Star Arm 102.

  • Links. github.com/Seeed-Projects/reBot-DevArm · wiki.seeedstudio.com/robotics_page/ · LeRobot docs page rebot_b601

  • GitHub stats. 3.9k stars, 388 forks, 173 commits [gh].

  • License. Hardware CERN-OHL-W-2.0; software Apache-2.0. OSHWA certified (CN000024). Per the GitHub page summary (not cross-checked), the hardware license moved from "CC BY-SA NC" to CERN-OHL-W in May 2026.

  • What it is. Six DoF plus a gripper, with a sheet-metal and 3D-printed structure and a bill of materials down to individual screws.

    B601-DMB601-RS
    MotorsDamiao 4340P (base), 4310 (wrists)RobStride rs-06 (base), rs-00 (wrists)
    Supply24 V48 V
    Payload1.5 kg2.5 kg
    Max reach767 mm754 mm
    Weight≈4.5 kg≈6.7 kg
    Repeatability<0.2 mm<0.2 mm
    • Recommended workspace is 70% of reach.
    • Software: MotorBridge Python SDK, ROS 1/2 (MoveIt 2 for RS), Pinocchio, Isaac Sim USD files plus an NVIDIA "Sim-to-Real VLA" course, a browser MuJoCo twin, and the Menagerie model seeed_rebot_devarm.
  • Price (Seeed store, B601-DM unassembled, 2026-10-02): motor kit $994.80 + structure $202.80 + gripper $238.80 = $1,436.40, plus $64.99 for the power kit. The Star Arm 102 leader is $240.

  • Adoption. 3.9k stars in about six months and LeRobot core integration. Hugging Face dataset counts were not fetched.

  • Benchmarks. The repo has a performance-testing document that I did not read.

  • Trade-offs.

    • Strengths: 1.5–2.5 kg payload and about 0.75 m reach for roughly $1.5–2k; a commercial-friendly hardware license; Seeed distribution and NVIDIA tie-ins.
    • Weaknesses: 6-DoF rather than 7; new, with few datasets; the leader is a servo design, so teleop is not bilateral; the RS variant needs a 48 V high-current supply.
    • It sits between SO-101 and OpenArm/Piper.
  • Sources. [idx] Seeed-Projects/reBot-DevArm README and hardware/ BOM files · [gh] https://github.com/Seeed-Projects/reBot-DevArm · https://www.seeedstudio.com/reBot-Arm-B601-DM-Bundle.html · [idx] lerobot docs/source/rebot_b601.mdx

Koch v1.1 and Moss v1 — Dynamixel- and STS3215-based predecessors of SO-100

  • Status.
    • Koch v1.1: DORMANT as a hardware design. Its README still links a 2024 LeRobot example path that no longer exists. It is still supported in LeRobot main (koch_follower/koch_leader are present).
    • Moss v1: DISCONTINUED (deprecated). Its README says "we now recommend building or purchasing the SO100 … it will be deprecated", and there is no moss type in LeRobot main.
  • Who. Alexander Koch designed the original low-cost arm. Jess Moss (Hugging Face) made Koch v1.1 and Moss.
  • Links. github.com/jess-moss/koch-v1-1 · github.com/jess-moss/moss-robot-arms
  • Stats. koch-v1-1: 720 stars, 66 forks. moss-robot-arms: 88 stars, 6 forks [gh].
  • License. Apache-2.0 for both (© 2024 Hugging Face) [idx].
  • What it is.
    • Koch: 5 DoF plus gripper. Leader: six Dynamixel XL330-M077-T (5 V). Follower: two XL430-W250-T and four XL330-M288-T (12 V). US bill of materials: $199 (leader) and $278 (follower).
    • Moss: six STS3215 per arm; $159 for one arm, $288 for a pair.
    • Menagerie has a low_cost_robot_arm model.
  • Adoption. Hugging Face datasets: 58 tagged koch, 957 with "koch" in the name. HF's 2025 blog said "most contributions are to So100 and Koch".
  • Trade-offs. Better Dynamixel encoders, but about twice the cost of SO-101, and superseded by it for new builds. LeKiwi has a Dynamixel/Koch variant.
  • Sources. [idx] both READMEs and LICENSE files · [gh] https://github.com/jess-moss · https://huggingface.co/datasets?other=koch, ?search=koch

PAROL6 (Source Robotics) — an industrial-style 3D-printed stepper arm

  • Status: ACTIVE (probable; indirect evidence). The Python API ships wheels for CPython 3.14 (released October 2025), and there is a new web commander. I could not verify 2026 commit dates.
  • Who. Petar Crnjak (GitHub PCrnjak) / Source Robotics, which sells kits and assembled arms. The repo is now under source-robotics.
  • Links. github.com/PCrnjak/PAROL6-Desktop-robot-arm · PAROL6-python-API · PAROL6-ROS2-MOVEIT · source-robotics.github.io/PAROL-docs
  • Stats. 3.3k stars, 397 forks, 166 commits [gh].
  • License. GPL-3.0.
  • What it is (self-reported docs):
    • Six rotary joints; 400 mm reach with the standard gripper.
    • Payload "1 kg near the base, 0.5 kg across the full workspace"; repeatability 0.1 mm.
    • 5.5 kg; 40 W; steppers with planetary and belt reduction; open-loop with limit-switch homing.
    • One CAN bus; USB connection to the PC.
    • Joint ranges: 250°, 141°, 180°, 212°, 180°, and continuous on J6.
    • No brakes: the arm falls on power loss.
    • ROS 2/MoveIt simulation is available. Not in LeRobot core or its third-party list.
  • Benchmarks (thermal tests in the docs): joints hold at 48–61 °C over 5 hours and reach 52–73 °C after 2 hours of motion. The docs warn that you "must reduce the current" for long sessions or risk destroying the PETG robot.
  • Trade-offs. Industrial kinematics and a GUI at low cost, but not backdrivable, thermally limited and copyleft. It is a poorer fit for imitation-learning data than SO-101 or OpenArm.
  • Sources. [gh] repo page · [idx] README · [idx] source-robotics/PAROL-docs docs/page2_2.md and page6.md · [idx] PAROL6-python-API pyproject.toml

Annin Robotics AR4 (MK3 / MK4) — a machined-aluminium stepper arm sold as kits

  • Status: ACTIVE. The site lists AR4 MK3 kits (combo, aluminium parts, hardware, electrical) and references an "AR4 MK4 manual released February 2026". The site was last modified 2026-06-09.
  • Who. Annin Robotics.
  • Links. anninrobotics.com · community ROS 2 driver ar4_ros_driver (index.ros.org; seen in search only)
  • Specs (RoboDK): 6 axes, 600 mm reach, 1.0 kg payload, 0.2 mm repeatability, 10 kg.
  • Software. Python, ROS 2 drivers, Arduino, Modbus, G-code.
  • License. A custom "license agreement" on the site, not a standard open license, so treat it as partially open.
  • Adoption. GELLO includes a community AR4 leader design. Prices are not shown on the landing page.
  • Trade-offs. Rigid and long-reach, but stepper-based (not backdrivable), heavy, under a custom license, with almost no LeRobot presence.
  • Sources. https://www.anninrobotics.com/ · https://robodk.com/robot/Annin-Robotics/AR4 · [idx] gello_mechanical README

Arctos — a 3D-printed 6-axis stepper arm sold as kits

Thor (AngelLM) — a 3D-printed G-code arm

  • Status: DORMANT (likely). The latest dated design files in the repo are 2025-04-17 [idx].
  • Stats. 1.6k stars, 271 forks, 354 commits [gh]. License: CC-BY-SA-4.0.
  • What it is. Six DoF in a yaw-roll-roll-yaw-roll-yaw layout; 625 mm reach; 750 g payload; steppers with 3D-printed gears and GT2 belts; GRBL/RRF firmware; can integrate with ROS 2; hardware from under €350.
  • Trade-offs. A very cheap and educational arm, not built for learning data.
  • Sources. [gh] https://github.com/AngelLM/Thor · [idx] grep of the repo

BCN3D Moveo — DORMANT

An educational 3D-printed arm from BCN3D and the Catalan Department of Education. 1.9k stars, 561 forks, 11 commits, MIT. No modern stack.

Source: [gh] https://github.com/BCN3D/BCN3D-Moveo

Dummy (Peng Zhihui) — DORMANT, not legally open

  • 15.6k stars, 3.2k forks, 34 commits, and no license specified, so all rights are reserved by default.
  • Six joints using size 20 and 42 steppers with harmonic reducers, CAN, and custom REF control boards. The Unity3D desktop app is not open source.
  • The author later co-founded AgiBot.

Source: [gh] https://github.com/peng-zhihui/Dummy-Robot

ROBOTIS OpenMANIPULATOR-X / OMX / OMY — Dynamixel arms with factory calibration

  • Status: ACTIVE.
    • OMX is in LeRobot core (omx_follower/omx_leader), "shipped preconfigured from the factory … no additional motor setup or calibration".
    • OMY is in MuJoCo Menagerie (robotis_omy_3m), and an OMY-L100 teleoperator plugin exists for LeRobot.
  • Who. ROBOTIS (Korea).
  • Links. github.com/ROBOTIS-GIT/open_manipulator (ROS 2 packages for OMY and OpenMANIPULATOR-X) · physical_ai_tools · robotis_mujoco_menagerie · huggingface.co/ROBOTIS
  • What it is.
    • OMX: 5 DoF plus gripper, the entry tier.
    • OMY: 6 DoF plus gripper, DYNAMIXEL-Y actuators, "collaborative robot style" with gravity compensation and self-collision detection.
    • OMY variants: L100 (leader), F3M (gripper plus RealSense D405), 3M, and AI3M (leader–follower set).
    • Reach, payload and price were not published on the pages I read.
  • License. Not verified this session. Stars: not fetched.
  • Trade-offs. Factory calibration solves the SO-101 calibration pain, and it is ROS 2 native with LeRobot support. It costs more, and openness of the hardware design is unverified.
  • Sources. https://ai.robotis.com/ · https://ai.robotis.com/omy/introduction_omy.html · [idx] open_manipulator README · [idx] lerobot omx.mdx

Waveshare RoArm-M3 (and M2) — an ESP32 Wi-Fi desktop arm

  • Status: ACTIVE. The wiki documentation is dated 2026-07-01.
  • Specs.
    • 5+1 DoF with a workspace of up to 1 m in diameter.
    • 0.2 kg payload at 0.5 m; 0.088° repositioning accuracy from 12-bit magnetic encoders.
    • ST3215 servos (plastic M3-S, metal M3-Pro).
    • ESP32-WROOM-32; Wi-Fi AP/STA plus USB serial with JSON commands; web UI.
    • ROS 2 and MoveIt 2 tutorials; claims LeRobot support.
  • Not verified. Price, license and the M2 specs.
  • Trade-offs. Cheap and wireless, but the payload is very low.
  • Source. https://www.waveshare.com/wiki/RoArm-M3

Reachy 2 arm (Pollen Robotics, now part of Hugging Face)

  • Status: SLOWING. reachy2-sdk was last pushed 2025-11-26, with v1.0.15 on 2025-10-29 [pkg]. It is still marketed (the site footer reads © 2026) and supported in LeRobot core (reachy2 robot and teleoperator).
  • Who. Pollen Robotics (France), acquired by Hugging Face (see the HF blog post on the acquisition).
  • What it is.
    • A humanoid upper body with 7-DoF arms, about 3 kg payload per arm, using Orbita 2D/3D parallel actuators.
    • A VR teleop app, not included in LeRobot.
    • Optional omnidirectional base with LiDAR.
    • ROS 2 Humble and a Python SDK; Apache-2.0 software; Gazebo/MuJoCo Docker simulation.
    • Price is not listed.
  • Stats. reachy2-sdk: 71 stars, 2.5k downloads/month [pkg].
  • Trade-offs. Polished, with turnkey VR teleop, but expensive and with a small dataset community.
  • Sources. https://pollen-robotics.com/reachy-2/ · pkg reachy2-sdk · [idx] lerobot reachy2.mdx

HopeJR arm (TheRobotStudio × Hugging Face)

  • Status: ACTIVE. hope_jr robot types plus the Homunculus arm and glove teleoperators are in LeRobot main.
  • Who. The Robot Studio (the humanoid design is credited to Rob Knight) and HF (the arm work is credited to Martino Russi; the Homunculus glove to "NepYope & Hugging Face").
  • What it is. A dexterous arm and hand driven by an exoskeleton with 1:1 joint mapping, needing no soldering. The humanoid version uses Robstride 02 motors.
  • Price. "$3,000 humanoid" according to a Notebookcheck headline (May 2025; search listing only).
  • Stats. 819 stars [gh].
  • License. No root LICENSE file was found, so the license is unclear.
  • Sources. [idx] TheRobotStudio/HOPEJr README · [idx] lerobot hope_jr.mdx and teleoperators/homunculus · [gh] TheRobotStudio org page

Berkeley Blue — DORMANT (discontinued in practice)

blue_core needs Ubuntu 16.04/18.04 with ROS Kinetic or Melodic; there is no modern stack.

Source: [idx] github.com/berkeleyopenarms/blue_core

AM-ARM200 (Li Yiteng, 2026) — an SO-class upgrade

  • Status: ACTIVE (README update dated 2026-05-29).
  • What it is. 6+1 DoF, 1 kg payload, 52 cm reach. Uses STS3215 servos plus three STS3095 servos (12 V). Costs about $240 for the follower, $140 for the leader and $380 for the pair. About 20 hours of printing and about 60 minutes of assembly.
  • LeRobot. Supported through the lerobot_alohamini fork or a small patch to upstream.
  • License. Apache-2.0. Stars: 76 [gh].
  • Sources. [idx] liyiteng/AM-ARM README · [gh] https://github.com/liyiteng

B. Bimanual and mobile-manipulation rigs

ALOHA / ALOHA 2 (with Trossen's commercial continuation)

  • Status.
    • The original repo is DORMANT: it targets ROS 1 Noetic on Ubuntu 18.04/20.04, and ROS 2 is still listed as "ongoing testing".
    • The ALOHA 2 hardware (2024) has not changed.
    • The ecosystem is ACTIVE through Trossen (lerobot_trossen updated 2026-07-24; trossen_arm 2026-08-11; trossen_sdk 2026-08-25) and through Menagerie (aloha model).
  • Who. ALOHA: Tony Z. Zhao and co-authors, including Chelsea Finn. ALOHA 2: 23 researchers from Google DeepMind, Stanford and Hoku Labs.
  • Links. github.com/tonyzhaozh/aloha (ALOHA 2 hardware under /aloha2) · tonyzhaozh/act · aloha-2.github.io · arXiv 2304.13705
  • Stats. tonyzhaozh/aloha: 2.3k stars, 337 forks. tonyzhaozh/act: 2.1k stars, 406 forks [gh].
  • License. MIT (software).
  • What it is.
    • Four Dynamixel arms: Interbotix/Trossen WidowX 250 leaders and ViperX 300 followers; Menagerie has trossen_wx250s and trossen_vx300s.
    • Four cameras; HDF5 episodes; 50 Hz; a "$20k budget".
    • Known pain points from the README: needs at least 6 USB 3 ports; machines struggle to connect all four arms stably, especially through hubs; device ports rebind (fixed with udev rules); forward-kinematics computation delays teleop unless you patch Interbotix code.
    • ALOHA 2 adds low-friction rail grippers on leader and follower, passive gravity compensation, a 20×20 mm extrusion frame and RealSense D405 cameras.
  • Adoption. Hugging Face datasets: 155 tagged aloha; 945 with "aloha" in the name, including simulation datasets.
  • Benchmarks. ACT with 50 demonstrations: slot battery 92%, open lid 84%, prepare tape 64%, put on shoe 96%. Some tasks reach 80–90% with 10 minutes of demonstrations.
  • Trade-offs. A proven recipe with many datasets, but $20k+, ROS 1 legacy and lots of USB devices. Trossen's 2026 lineup (see section C) is the supported path.
  • Sources. [gh] https://github.com/tonyzhaozh · [idx] aloha README and LICENSE · https://tonyzhaozh.github.io/aloha/ · https://aloha-2.github.io/ · [gh] https://github.com/TrossenRobotics · https://huggingface.co/datasets?other=aloha, ?search=aloha

Mobile ALOHA — DORMANT

  • 4.5k stars, 733 forks, MIT, 73 commits; ROS 1 Noetic; AgileX Tracer base; three cameras [gh].
  • Stanford (Zipeng Fu, Tony Zhao, Chelsea Finn); arXiv 2401.02117.
  • Co-training "can increase success rates by up to 90%" with 50 demonstrations per task.
  • The roughly $32k total cost comes from the paper and was not re-verified this session.
  • Sources. [gh] https://github.com/MarkFzp/mobile-aloha · https://mobile-aloha.github.io/

XLeRobot — a $660 dual-arm household mobile robot

  • Status: ACTIVE. The repo's docs/community_usecases/data.json was collected 2026-09-18 and checked 2026-09-24. Earlier news: browser MuJoCo plus 3D Gaussian splatting demo (2026-01-29); dual-wheel base (2025-12-02).
  • Who. Gaotian (Vector) Wang (Rice University RobotPi Lab), with the MakerMods team and others. Described as non-profit.
  • Links. github.com/Vector-Wangel/XLeRobot · xlerobot.readthedocs.io
  • Stats. 5.5k stars, 610 forks, 1,502 commits; release 0.3.0 on 2025-08-30 [gh].
  • License. Apache-2.0.
  • What it is.
    • Two SO-101 arms on a LeKiwi base, a 300 Wh Anker battery, two wrist RGB cameras and a head depth camera on a 2-DoF neck.
    • Self-sourced cost about $660 / €680 / ¥3,999. Kit $579 worldwide or ¥3,699 on Taobao (WowRobo, excluding battery and IKEA cart).
    • Assembly under 4 hours.
    • Teleop by keyboard, Xbox controller, Joy-Con or Quest 3 VR; simulation includes an RL environment.
  • Adoption. 43 catalogued community uses, including papers at ACL 2026, AAAI 2026 and HRI 2026, and integrations with IsaacLab, Gazebo, ROS 2, LiveKit and WebXR. 257 Hugging Face datasets have "xlerobot" in the name.
  • Trade-offs. The cheapest bimanual mobile platform, but it inherits SO-101's payload and calibration limits.
  • Sources. [gh] repo page · [idx] README and docs/community_usecases/data.json · https://huggingface.co/datasets?search=xlerobot

LeKiwi (SIGRobotics UIUC × Hugging Face) — a holonomic base for SO-101

  • Status. The v1 hardware design is stable, with no dated 2026 revision found. Software is ACTIVE through LeRobot core (lekiwi). LeIsaac added LeKiwi simulation on 2025-12-19.
  • Who. CAD by Manav Chandaka, Bhargav Chandaka and Pepijn Kooijmans (HF).
  • Stats. 1.4k stars, 155 forks, 103 commits [gh]. License: Apache-2.0.
  • What it is.
    • Three omni-wheel kiwi drive, Raspberry Pi 5 streaming to a laptop, SO-101 arm, workspace and wrist cameras.
    • 5 V (power bank) or 12 V (Li-ion) versions; URDF provided.
    • A Dynamixel variant uses Koch arms.
    • BOM.md section totals range from $31 to $212; the arm is extra.
  • Adoption. Hugging Face datasets: 76 tagged lekiwi, 778 with "lekiwi" in the name.
  • Sources. [gh] repo page · [idx] README and BOM.md · HF dataset searches (?other=lekiwi, ?search=lekiwi)

AlohaMini / AlohaMini2 (Li Yiteng, 2026) — a dual-arm mobile robot with a lift for under $1,000

  • Status: ACTIVE. AlohaMini2 was released 2026-06-06; a Pi 0.5 openpi guide was added 2026-02-26.
  • What it is.
    • 70 kg base load, 30 kg lift, two AM-ARM200 arms (1 kg payload, 6+1 DoF, 52 cm reach), five cameras.
    • Self-build bill of materials about $950; about 120 minutes of assembly; prints on a consumer Bambu P2S.
  • License and stats. Apache-2.0; 1.1k stars, 134 forks [gh].
  • Sources. [idx] liyiteng/AlohaMini README · [gh] https://github.com/liyiteng

TidyBot++ (Stanford/Princeton) — an open holonomic base built around a Kinova arm

  • Status: DORMANT. Only 5 commits. A Menagerie model (stanford_tidybot) exists.
  • What it is. A Kinova Gen3 arm on a powered-caster holonomic base, teleoperated through a phone web app. Policy inference takes 115 ms on an RTX 4080 laptop GPU; control runs at 10 Hz.
  • License and stats. MIT; 609 stars [gh].
  • Source. [gh] https://github.com/jimmyyhwu/tidybot2

ROBOTIS AI Worker (brief)


C. Closed arms commonly used with open stacks (comparators)

  • AgileX Piper.

    • $1,999; 6-DoF; 626 mm reach; 1.5 kg payload; 0.1 mm repeatability; 4.2 kg.
    • piper_sdk (MIT): 404 stars, v0.6.2 on 2026-07-31.
    • Two LeRobot plugins and a Menagerie model. Hugging Face: 114 datasets tagged piper; the name search returned 823, but it is contaminated by "Piper" text-to-speech voice datasets.
    • Sources: https://global.agilex.ai/products/piper · pkg piper_sdk
  • Trossen WidowX 2026 lineup (pricing marked "proposed and pending final approval"):

    ModelPriceReachPayload
    Core Follower$2,995 ($3,495 with camera)700 mm2 kg
    Pro$3,495700 mm4 kg
    Heavy$4,9951.5 m8 kg
    Core Leader$3,099.99——
    • "iNerve" controller at ">600 Hz", industrial Ethernet, open-source Trossen SDK.
    • The lineup also includes the Rivet mobile base, Workbench, the "Glide + Cockpit" teleop system, the TRumi gripper and the TOTL workstation.
    • The legacy WidowX 250 S and ViperX 300 S are the ALOHA arms.
    • Sources: https://www.trossenrobotics.com/widowx · /2026-product-lineup
  • I2RT YAM family (YAM, YAM Pro, YAM Ultra, YAM Ultra 2, Big YAM).

    • All 6-DoF; arm masses 4.29–5.31 kg; Damiao motors over CAN.
    • MIT-licensed Python API with MuJoCo gravity compensation and bilateral "minimum_gello" teleop.
    • GELLO's recommended robot; added to Menagerie 2025-05-19; LeRobot plugins (lerobot_yam, an active YAM leader).
    • Price not verified.
    • Sources: [idx] i2rt-robotics/i2rt README and LICENSE
  • uFactory Lite 6. $3,349 ("from $2,999"); 440 mm; 600 g; ±0.5 mm; 8 kg. The xArm 6 has 700 mm reach and 5 kg payload; price not shown. xarm-python-sdk (BSD-3) v1.18.4 on 2026-05-21. Source: https://www.ufactory.cc/lite-6-collaborative-robot/

  • Unitree Z1 Air / Pro. 740 mm reach; 2 kg / ≥3 kg payload; ~0.1 mm; 4.3 / 4.5 kg; Ethernet; force control; price not listed. The Unitree D1 was not verified. Source: https://www.unitree.com/z1

  • Elephant myCobot. pymycobot (MIT): 211 stars, v4.0.7 on 2026-09-05. Price not verified.

  • Franka FR3, Kinova Gen3, UR5e. Prices are quote-based and not published. All three have Menagerie models.

    • FR3 has an official Franka-maintained GELLO ROS 2 integration and a pre-assembled "Franka GELLO Duo".
    • There are LeRobot extensions for FR3 (LeFranX) and UR5e.
    • TidyBot++ uses Kinova.
  • ARX5 / L5. A LeRobot plugin and a Menagerie model exist; price not verified.


D. Teleoperation and data-collection hardware and software

LeRobot teleop tooling (Hugging Face)

  • Status: ACTIVE. v0.6.0 (2026-07-06) and v0.6.1 (2026-08-03).
  • Teleoperators in core: leader arms (SO, Koch, OMX, OpenArm, OpenArm Mini, reBot 102, Reachy 2), Homunculus arm and glove, gamepad, keyboard, phone, and Unitree G1.
  • Phone teleop: iOS through HEBI Mobile I/O (ARKit); Android through the WebXR teleop package.
  • Isaac Teleop example: an XR controller drives SO-101 over CloudXR.
  • Third-party plugins load automatically (packages prefixed lerobot_teleoperator_*): Quest WebXR, PICO 4, SpaceMouse, Haply Inverse3, omega.7, GELLO, LiveKit remote teleop, and more.
  • v0.6.0 features:
    • lerobot-rollout with DAgger: take over with the leader arm or a USB foot pedal; correction frames are tagged intervention.
    • "sentry" and "highlight" recording modes.
    • Foxglove streaming.
    • LeLab browser GUI (SO-101 only so far).
    • HF Jobs cloud training.
    • Depth support and a VLM language-annotation pipeline.
  • Rate. The docs' recording examples use 30 fps.
  • Sources. [idx] lerobot docs (third_party_robots.mdx, phone_teleop.mdx, isaac_teleop.mdx, hardware_guide.mdx) and src tree · [idx] huggingface/blog lerobot-release-v060.md

GELLO (UC Berkeley) — kinematically matched 3D-printed Dynamixel leaders

  • Status: ACTIVE. ROS 2 releases: ros2-v2.0.0 (2026-02-16), v2.1.0 (2026-04-08), v2.2.0 (2026-08-31).
  • Who. Philipp Wu, Yide Shentu, Zhongke Yi, Xingyu Lin, Pieter Abbeel. Franka Robotics GmbH contributes the FR3 ROS 2 code.
  • Links. github.com/wuphilipp/gello_software · gello_mechanical · arXiv 2309.13037
  • Stats. gello_software: 538 stars, 178 forks, 201 commits [gh]. gello_mechanical: stars not fetched.
  • License. MIT for software and mechanical. CAD files are available by email request only.
  • Robots supported: Franka FER, FR3 / FR3 Duo, UR5, xArm7, Lite6, I2RT YAM (recommended), AR4. Also offers FACTR gravity compensation and bimanual use.
  • Rate. The ROS 2 publisher was hard-coded to 25 Hz at 57,600 baud until v2.2.0 made both configurable.
  • Cost. Reported as under $300 per arm on the project site (not re-verified; site fetch failed).
  • Benchmarks. The paper's user study found GELLO "more reliable and efficient" than VR controllers and SpaceMouse.
  • Trade-offs. Cheap, intuitive joint-space teleop, but one design per target arm and no force feedback.
  • Sources. [gh] repo page · [idx] README, ros2/CHANGELOG.md, gello_mechanical README and LICENSE · https://arxiv.org/abs/2309.13037

NVIDIA Isaac Teleop (new)

  • Status: ACTIVE. PyPI isaacteleop 1.4.145 on 2026-09-17; 150 versions [pkg].
  • Stats. 327 stars, 69 forks, 686 commits; Apache-2.0 [gh].
  • Positioning. Described as "the unified framework for high-fidelity ego-centric and robotics data collection".
  • Inputs: XR controllers, hand tracking, full-body tracking, Manus gloves, foot pedals. Gamepad, GELLO and Haply are listed as upcoming.
  • Uses: HOMIE and SONIC loco-manipulation, egocentric collection without a robot, and Isaac Lab 3.0.
  • Transport: CloudXR/OpenXR, including an Apple Vision Pro profile. The first run requires accepting a EULA, and launch blocks for about 30 seconds.
  • Sources. [gh] https://github.com/NVIDIA/IsaacTeleop · [idx] lerobot isaac_teleop.mdx

Unitree xr_teleoperate (Vision Pro, PICO 4 Ultra, Quest 3)

  • Status: ACTIVE. v1.6 (2026-07-29) added the H2 and R1 robots and BrainCo hands. v1.4 (2025-11-21) moved to WebRTC head-camera streaming through teleimager and televuer v3.0, and added a passthrough mode.
  • Robots: G1 (29/23-DoF), H1, H1_2, H2, R1 (5-DoF or 7-DoF arm); hands Dex1-1, Dex3-1, Inspire, BrainCo.
  • Stats and license. 1.7k stars, 340 forks, 199 commits [gh]; Apache-2.0.
  • Latency. No figures published.
  • Sources. [gh] repo page · [idx] LICENSE and CHANGELOG.md

Open-TeleVision (UCSD / MIT, CoRL 2024) — DORMANT

  • Immersive stereo active vision on Apple Vision Pro or Quest 3 with a ZED camera; Isaac Gym simulation; ACT.
  • 1.3k stars, 135 forks, 22 commits; Apache-2.0. Unitree's televuer continues its lineage.
  • Sources: [gh] repo page · [idx] LICENSE

Vuer (Ge Yang, MIT/UCSD) — WebXR visualization and teleop backbone

  • Status: SLOWING. Last release v0.1.6 on 2026-03-20.
  • 405 stars, 25 forks, 816 commits [gh]; MIT [pkg]. Runs in the browser through Pyodide.
  • Sources: [gh] repo page · pkg vuer

Bunny-VisionPro — DORMANT

AnyTeleop / dex-retargeting (UCSD × NVIDIA) — DORMANT, but heavily used

  • Only the retargeting component is open (dex-retargeting).
  • 1.2k stars, 150 forks; MIT; v0.5.0 on 2025-04-05; last push 2025-08-21; 121k PyPI downloads/month [pkg].
  • Source: pkg dex-retargeting

VisionProTeleop (MIT Improbable AI) — Apple Vision Pro app plus avp_stream

  • Status: SLOWING. avp_stream 2.51 on 2025-12-28 [pkg].
  • 825 stars, 65 forks; MIT [gh].
  • Latency (self-reported): under 100 ms wireless at up to 720p; about 50 ms wired with 4K stereo.
  • Features: 27-joint hand skeleton; MuJoCo/Isaac Lab rendering in AR; egocentric recording with a USB camera; cloud recordings with optional public dataset sharing; WebRTC/TURN for remote operation.
  • Source: [gh] https://github.com/Improbable-AI/VisionProTeleop

Quest and WebXR apps

  • phosphobot:
    • SLOWING. Last PyPI release 0.3.134 on 2025-10-22; 150 versions.
    • MIT; 395 stars, 84 forks, 2,057 commits.
    • Y Combinator W24.
    • A Meta Quest app; supports SO-100/101, plus Koch, WX-250, Piper and Go2 in beta.
    • Paid "phospho PRO" at €35/month includes cloud training and VR teleop.
    • Sells "Custom: Datasets, AI models, evaluations" — a direct competitor for a data startup.
    • Self-reported "2,000+ robots" and "1,000+ models".
  • trossen_vr: Meta Quest app plus C++/Python SDK with dual-arm deadman switches; updated 2026-08-04 [gh].
  • SpesRobotics teleop: WebXR teleop from a phone or headset; Apache; 282 stars; v0.1.5 on 2026-04-19, last pushed 2026-09-18 [pkg]. LeRobot uses it for Android phone teleop.
  • OpenArm VR (Quest 3, about 20 FPS, simulation only per the docs) and dora-openarm-webxr (updated 2026-09-29).
  • Sources: [gh] https://github.com/phospho-app/phosphobot · https://phospho.ai/ · [idx] TrossenRobotics/trossen_vr README · pkg teleop

HOMIE (Shanghai AI Lab / InternRobotics) — exoskeleton cockpit for Unitree G1

  • Status: DORMANT. Released February 2025; 16 commits.
  • What it is. Isomorphic exoskeleton arms, 15-DoF motion-sensing gloves and a locomotion pedal. Hardware costs about $0.5k (self-reported).
  • License: CC BY-NC-SA 4.0, and the README says "strictly forbidden to use HOMIE for any commercial purposes". Some hardware files require a form with an institutional affiliation.
  • Stats. 619 stars, 49 forks [gh].
  • Source: [gh] https://github.com/InternRobotics/OpenHomie

AirExo / AirExo-2 (SJTU) — exoskeleton for teleop and robot-free in-the-wild collection

  • Status: DORMANT. Initial release 2025-05-15; sample data added 2025-06-01.
  • What it is. Paired with two Flexiv Rizon 4 arms and Robotiq 2F-85 grippers; uses the RISE-2 policy; arXiv 2503.03081.
  • License: CC BY-NC-SA 4.0.
  • Source: [idx] AirExo/AirExo-2 README and LICENSE

ACE (UCSD, CoRL 2024) — cross-platform visual exoskeleton

  • Status: DORMANT.
  • Robots: xArm with Ability Hand, Franka with gripper, Unitree H1 with Inspire hand, GR-1 with gripper. Separate hardware repo (ACETeleop/ACE_hardware). arXiv 2408.11805.
  • License. Apache-2.0.
  • Source: [idx] ACETeleop/ACETeleop README and LICENSE

UMI-style handheld grippers

  • UMI (Stanford/Columbia/TRI): a hand-held gripper with ORB-SLAM3. Menagerie has umi_gripper. Activity not verified.
  • Trossen TRumi (MIT; updated 2026-07-10): an ORB-SLAM3 dataset pipeline.
  • Sources: [idx] real-stanford/universal_manipulation_interface README · [idx] TrossenRobotics/trumi README

1. Also notable (brief)

  • LeIsaac (Lightwheel) — ACTIVE. SO-101 and LeKiwi teleop in Isaac Lab; LeRobot recorder; a datagen state machine (2026-03-10); remote teleop (2026-04-14); GR00T N1.6; the official LeRobot "EnvHub" imitation-learning playground. Apache-2.0. Lightwheel is a simulation-data company, so a potential competitor. [idx] github.com/LightwheelAI/leisaac
  • MuJoCo Menagerie — ACTIVE (pip package added 2026-09-01). Includes SO-ARM100/SO-101, reBot, OMY, WidowX AI, VX300s, WX250s, Piper, YAM, Lite6, xArm7, Z1, FR3, Gen3, UR5e, ALOHA and the UMI gripper. No OpenArm (Enactic maintains its own). [idx] google-deepmind/mujoco_menagerie
  • TRLC-DK1 (The Robot Learning Company, Germany) — a LeRobot plugin bimanual development kit (follower v0.3.0 STEP files); the README says "All rights reserved", so the license is unclear. [idx]
  • Nextis AIRA — a 3D-printable 6+1-DoF arm with motors up to 35 Nm peak; Apache-2.0; "early access". [idx]
  • Hiwonder NexArm; FashionStar StarAI Cello / Viola / Star Arm 102; HEXFELLOW; Lebai; ugo Pro; ARX5 — LeRobot third-party plugins; status not verified. [idx] lerobot third_party_robots.mdx
  • HF LeLab — a browser GUI for LeRobot; SO-101 only. [idx] v0.6.0 blog

2. Comparison tables

Open arms (n/p = not published; n/v = not verified this session)

ArmDoFReachPayloadRepeatabilityPrice (USD)License (HW / SW)Status
SO-1015+1n/pn/pn/p (calibration error ±50–200 ticks)$122 follower BOM; $230 pair; kits $100–$500Apache-2.0ACTIVE
Koch v1.15+1n/pn/pn/p$278 follower + $199 leader (BOM)Apache-2.0DORMANT (HW); supported
Moss v15+1n/pn/pn/p$159 per armApache-2.0DISCONTINUED
AM-ARM2006+1520 mm1 kgn/p~$240 / ~$380 pairApache-2.0ACTIVE
reBot B601-DM6+1767 mm1.5 kg<0.2 mm~$1,436 kit (+$240 leader)CERN-OHL-W-2.0 / Apache-2.0ACTIVE
reBot B601-RS6+1754 mm2.5 kg<0.2 mmn/vCERN-OHL-W-2.0 / Apache-2.0ACTIVE
OpenArm 2.07+1 per arm606 mm4.1 kg nominal / 6.0 kg peakn/p$5.6k–$7.3k bimanual; KER $2.4k–$2.6kCERN-OHL-S-2.0 / Apache-2.0ACTIVE
PAROL66400 mm1 kg near base / 0.5 kg0.1 mmn/vGPL-3.0ACTIVE (probable)
AR4 (MK3/MK4)6600 mm1.0 kg0.2 mmn/vcustomACTIVE
Thor6625 mm0.75 kgn/p<€350 partsCC-BY-SA-4.0DORMANT (likely)
Arctos V26n/vn/vn/v$452 non-printed kitn/vACTIVE (commercial)
RoArm-M35+1~0.5 m0.2 kg at 0.5 m0.088°n/vn/vACTIVE
OMY / OMX6+1 / 5+1n/pn/pn/pn/vn/vACTIVE
Reachy 2 arm7n/p~3 kgn/pn/pApache-2.0 (SW)SLOWING
Moveo / Dummy / Blue—————MIT / none / n/vDORMANT

Closed comparators

ArmDoFReachPayloadRepeatabilityPrice
AgileX Piper6626 mm1.5 kg0.1 mm$1,999
WidowX Core / Pro / Heavy (2026)n/p700 / 700 / 1,500 mm2 / 4 / 8 kgn/p$2,995 / $3,495 / $4,995 (proposed)
uFactory Lite 66440 mm0.6 kg±0.5 mm$3,349
Unitree Z1 Air / Pro6740 mm2 / ≥3 kg~0.1 mmn/p
YAM; Franka FR3; Kinova Gen3; UR5e————quote-based / n/v

Teleop systems

SystemTypeCostRobotsRate / latencyLicenseStatusStars
SO-101 leader (LeRobot)joint-matched leader arm~$108 incremental BOMSO-100/10130 fps recording typicalApache-2.0ACTIVE—
GELLOjoint-matched leader<$300 (reported)FR3, FER, UR5, xArm7, Lite6, YAM, AR4ROS 2 default 25 Hz (configurable since Aug 2026)MITACTIVE538
ALOHA leadersWidowX 250 leaderspart of $20k systemViperX 30050 HzMITDORMANT / Trossen ACTIVE2.3k
OpenArm leader / KER / Minibilateral QDD / motorless / servoKER $2,399–$2,599OpenArmfollower 1 kHz CAN-FDApache-2.0 + CERN-OHL-SACTIVE3.5k
Isaac TeleopXR, gloves, pedals, egocentricfree softwareSO-101 example, HOMIE/SONICn/pApache-2.0ACTIVE327
xr_teleoperateAVP / PICO / QuestfreeUnitree humanoids and handsWebRTC; n/pApache-2.0ACTIVE1.7k
VisionProTeleopAVP app + libraryfreeany (Python)<100 ms wireless, ~50 ms wiredMITSLOWING825
Open-TeleVisionAVP / Quest + stereofreesim, humanoidsn/pApache-2.0DORMANT1.3k
VuerWebXR toolkitfree—n/pMITSLOWING405
Bunny-VisionProAVPfreexArm7 + Ability Handn/pMITDORMANT354
dex-retargeting (AnyTeleop)hand retargetingfreedexterous handsn/pMITDORMANT1.2k
phosphobotQuest app + web UIfree; PRO €35/moSO arms and othersn/pMITSLOWING395
SpesRobotics teleopWebXR phone/headsetfreeURDF-based armsn/pApacheACTIVE282
HOMIEexoskeleton + gloves + pedal~$0.5kUnitree G1n/pCC BY-NC-SADORMANT619
AirExo-2exoskeleton, in-the-wildn/pFlexiv Rizon 4n/pCC BY-NC-SADORMANTn/v
ACEvisual exoskeletonn/pxArm, Franka, H1, GR-1n/pApache-2.0DORMANTn/v

3. Gaps and pain points (with evidence) — relevance to a data, simulation or environment business

  1. The open data commons is small and dirty.
    • LeRobot's curated community dataset has 791 datasets, 50,622 episodes and 251.5 hours. Curators started from 851 and dropped 15–20 datasets for missing videos and 10–15 for data-type issues, and reprocessed 16.
    • HF's blog lists empty or meaningless task strings ("Hold", "desc"), ambiguous camera keys (images.laptop used for both third-person and wrist views), 1-frame episodes, deleted parquet files, and inconsistent action/state dimensions "even for the same robot (so100)".
    • Hub metadata is unreliable: 9,432 datasets have "so101" in the name, but only 1,237 are tagged so101 and only 15 are tagged so101_follower. There is no first-class embodiment filter.
    • Opportunity: QA'd, consistently annotated, embodiment-normalised data; curation and search tooling. Note that HF is moving into annotation with LeRobot v0.6's VLM annotation pipeline.
  2. "Identical" arms are not identical.
    • SO-101 hand calibration error is ±50–200 ticks per joint.
    • There are two calibration zero conventions.
    • The LeRobot gripper mapping is missing from the URDF/MJCF; motor parameters are borrowed from another project; base collision meshes were removed.
    • OpenArm built a mechanical zero-position jig "eliminating assembly tolerances from the dataset".
    • Opportunity: per-unit calibration and system identification, calibrated digital twins, real-to-sim alignment for SO-101, OpenArm and reBot.
  3. Format fragmentation.
    • LeRobot v2.1 → v3.0 migration; OpenArm's own format (with converters); ALOHA HDF5; Unitree EpisodeWriter; AirExo, Isaac Teleop and VisionPro cloud recordings; GR00T modality.json.
    • Opportunity: conversion and validation layers, and a canonical schema.
  4. Teleop throughput and ergonomics.
    • Powered leaders cause fatigue in long sessions, which is why KER exists.
    • GELLO's ROS 2 default was 25 Hz until August 2026.
    • VR and phone teleop has no force feedback; Vision Pro streaming adds 50–100 ms; the OpenArm VR prototype runs at about 20 FPS in simulation only.
    • Bilateral teleop needs QDD arms (OpenArm, YAM).
    • Opportunity: low-fatigue, high-rate collection stations with operator playbooks and quality control.
  5. Robot-free and egocentric collection is being platformised by incumbents.
    • Isaac Teleop does egocentric collection without robots and supports Manus gloves.
    • VisionProTeleop offers egocentric recording and public sharing.
    • Trossen has TRumi; AirExo-2 does in-the-wild collection; phospho sells custom datasets; Lightwheel generates simulation data.
    • Implication: differentiation must come from data diversity, operations and quality, not from teleop software.
  6. Long-session reliability.
    • PAROL6 has thermal limits and no brakes.
    • STS3215 gears strip, which is why the calibration tool caps torque at 30%.
    • ALOHA has USB instability with four arms.
    • OpenArm's CAN stack is Linux-only; brltty hijacks serial ports (noted in the reBot docs); Koch's docs have gone stale.
    • Opportunity: fleet health monitoring and turnkey data cells.
  7. Real-world evaluation is not standardised.
    • The OpenArm Cell exists because cross-lab comparisons are meaningless.
    • LeRobot v0.6 unifies nine simulation benchmark families, but real-world evaluation is still bespoke.
    • OpenArm has no Menagerie model, and the SO-101 leader has no MJCF.
    • Opportunity: matched simulation environments plus standardised real-world evaluation cells or services.
  8. Licensing traps for a commercial data company.
    • Non-commercial: HOMIE, AirExo-2.
    • No license: Dummy. "All rights reserved": TRLC-DK1 README.
    • Copyleft or reciprocal: PAROL6 GPL-3.0, OpenArm CERN-OHL-S.
    • Gated: GELLO CAD on request; HOMIE behind an institutional form.
    • reBot's May 2026 switch from non-commercial to CERN-OHL-W shows vendors moving toward commercial terms.
  9. Capability split.
    • SO-class arms lack payload and force sensing, so contact-rich data needs $2k–$7k QDD or torque arms.
    • Datasets for those mid-tier arms are each tiny (39 tagged openarm; 114 tagged piper).
    • Opportunity: cross-embodiment mid-tier datasets with standardised grippers and wrist cameras.

4. GitHub repositories referenced (stars fetched 2026-10-02)

  • huggingface/lerobot — 27k [pkg]
  • TheRobotStudio/SO-ARM100 — 7.2k
  • TheRobotStudio/HOPEJr — 819
  • TheRobotStudio/V1.0-Dexhand — 555
  • jess-moss/koch-v1-1 — 720
  • jess-moss/moss-robot-arms — 88
  • AlexanderKoch-Koch/low_cost_robot — not fetched
  • enactic/openarm — 3,534
  • enactic/openarm_mujoco — 61
  • enactic/openarm_mjlab — 12
  • enactic/openarm_dataset — 11
  • enactic/openarm_control — 5
  • enactic/dora-openarm-webxr — 2
  • enactic/dora-openarm-data-collection — 4
  • enactic/openarm_hardware, openarm_can, openarm_description — not fetched
  • Seeed-Projects/reBot-DevArm — 3.9k
  • liyiteng/AM-ARM — 76
  • liyiteng/AlohaMini — 1.1k
  • liyiteng/lerobot_alohamini — 102
  • PCrnjak/PAROL6-Desktop-robot-arm — 3.3k
  • PCrnjak/PAROL6-python-API — not fetched
  • source-robotics/PAROL-docs — not fetched
  • AngelLM/Thor — 1.6k
  • BCN3D/BCN3D-Moveo — 1.9k
  • peng-zhihui/Dummy-Robot — 15.6k
  • ArctosRobotics/ROS — not fetched
  • berkeleyopenarms/blue_core — not fetched
  • ROBOTIS-GIT/open_manipulator — not fetched
  • ROBOTIS-GIT/ai_worker — 186
  • pollen-robotics/reachy2-sdk — 71 [pkg]
  • tonyzhaozh/aloha — 2.3k
  • tonyzhaozh/act — 2.1k
  • MarkFzp/mobile-aloha — 4.5k
  • Vector-Wangel/XLeRobot — 5.5k
  • SIGRobotics-UIUC/LeKiwi — 1.4k
  • jimmyyhwu/tidybot2 — 609
  • TrossenRobotics/trossen_sdk — 19
  • TrossenRobotics/trossen_arm — 14
  • TrossenRobotics/lerobot_trossen — 6
  • TrossenRobotics/trossen_arm_mujoco — 19
  • TrossenRobotics/trossen_ai_isaac — 13
  • TrossenRobotics/trumi — 1
  • TrossenRobotics/trossen_vr — 0
  • agilexrobotics/piper_sdk — 404 [pkg]
  • xarm-developer/xarm-python-sdk — 327 [pkg]
  • elephantrobotics/pymycobot — 211 [pkg]
  • i2rt-robotics/i2rt — not fetched
  • wuphilipp/gello_software — 538
  • wuphilipp/gello_mechanical — not fetched
  • NVIDIA/IsaacTeleop — 327
  • unitreerobotics/xr_teleoperate — 1.7k
  • OpenTeleVision/TeleVision — 1.3k
  • vuer-ai/vuer — 405
  • Dingry/BunnyVisionPro — 354
  • dexsuite/dex-retargeting — 1.2k [pkg]
  • Improbable-AI/VisionProTeleop — 825
  • phospho-app/phosphobot — 395
  • spesrobotics/teleop — 282 [pkg]
  • InternRobotics/OpenHomie — 619
  • AirExo/AirExo-2 — not fetched
  • ACETeleop/ACETeleop — not fetched
  • real-stanford/universal_manipulation_interface — not fetched
  • LightwheelAI/leisaac — not fetched (page fetch failed)
  • google-deepmind/mujoco_menagerie — not fetched
  • robot-learning-co/trlc-dk1 — not fetched
  • robertorobotics/Nextis-AIRA-3D — not fetched