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 humanoid whole-body control (WBC) — landscape as of 2026-10-02

Slice covered: RL locomotion frameworks, motion tracking and imitation, human-to-humanoid retargeting, teleop-to-WBC pipelines, model-based control libraries, plus the motion datasets these pipelines use.

Read this first: how the data was collected

• Pages I could not read. At the start of the session the WebFetch proxy rate-limited me (HTTP 429), and its instructions said not to re-fetch these pages:

  • github.com/NVlabs/GR00T-WholeBodyControl
  • github.com/amazon-far/holosoma
  • github.com/YanjieZe/GMR
  • github.com/HybridRobotics/whole_body_tracking
  • github.com/leggedrobotics/rsl_rl
  • pypi.org/project/mink
  • two humanoidsdaily.com articles (SONIC, holosoma)
  • arxiv.org/abs/2512.01996

  For each repo I used other pages instead: the /releases page (which shows stars and forks), README/LICENSE blob views, /issues, arXiv HTML and Hugging Face cards.

• Search ran out early. The session-wide WebSearch budget (200 calls, shared with other agents) was used up after about 25 of my searches. The fetch budget (800 per hour, shared) also ran out once. Some 2026 projects therefore appear only from search-result titles, and I mark them that way.

• Wrong years on GitHub release pages. The fetch summarizer gave wrong years for several repos (LocoMuJoCo, Pinocchio, Isaac Lab, robot_lab, aligator, whole_body_tracking). Where PyPI was available I used its dates. Where I corrected a year by inference, I say so.

• Star counts were fetched on 2026-10-02. GitHub rounds them (for example "2.4k"), and a few values were rendered as "3,000", which probably means "3k". Commit dates are not visible to the fetch tool; /commits, /tags and .atom pages are blocked by robots.txt and I did not work around that. So "status" relies on releases, PyPI, README news and paper dates.

Executive summary

1. Generalist trackers with released weights are now the frontier.
   • NVIDIA's SONIC (inside GR00T-WholeBodyControl) is the biggest open release in this area. Code is Apache-2.0 and weights use the NVIDIA Open Model License. Training code came out on 2026-04-10 and a v1.1 checkpoint on 2026-07-23.
   • SONIC was trained on 700 h of in-house mocap, which has not been released.
   • Smaller generalists:
     • GMT (checkpoint only)
     • TWIST/TWIST2 (full training code)
     • BFM-Zero (non-commercial license)
     • Humanoid-GPT (CVPR 2026; inference only)
2. Stacks are consolidating onto a few backends.
   • Isaac Gym repos from 2024–25 (H2O, HOVER, ASAP, HOMIE, PBHC, TWIST2) are mostly dormant, but many still depend on Isaac Gym, which needs Python 3.8.
   • Active 2026 stacks:
     • Isaac Lab 3.0 (PhysX plus Newton; v3.0.0-EA on 2026-09-16)
     • mjlab (Isaac Lab-style API on MuJoCo Warp; v1.6.0 on 2026-08-09)
     • Amazon FAR's holosoma (Isaac Gym, Isaac Sim and MuJoCo Warp; G1 and T1)
     • NVIDIA ProtoMotions3 (Isaac Gym, Isaac Lab 3, Newton, MuJoCo)
3. Retargeting is now its own component.
   • GMR: MIT license, 17 humanoids, real-time on CPU.
   • OmniRetarget: preserves contact with objects and terrain; its code lives in holosoma.
   • GMR's paper shows retargeting quality decides whether tracking policies learn at all: dance clips retargeted with PHC reached 0% success, versus 100% with GMR.
4. Unitree G1 is the default robot. Nearly every 2025–26 tracker is G1-first. Booster T1/K1 comes second.
5. Motion-data licensing is the binding constraint for commercial users.
   • AMASS forbids training models for commercial use and redistributing derivatives.
   • LAFAN1 is CC BY-NC-ND.
   • BONES-SEED (288 h) is gated under a custom license.
   • OmniRetarget and PHUMA both left LAFAN1-derived data out of their public releases because of licensing.
6. Model-based libraries are healthy. Pinocchio 4.1, Crocoddyl 3.2.1, TSID 1.10, Drake 1.57, Pink 4.4 and PlaCo 0.10 all released in 2026. They now mostly serve as IK, retargeting and hybrid-control components. The integrated MPC+WBC humanoid stacks (legged_control, OpenLoong-Dyn-Control) are discontinued or dormant.

A. Learning-based WBC: generalist trackers and full stacks

NVIDIA GR00T-WholeBodyControl / GEAR-SONIC — generalist humanoid motion-tracking model plus deployment and teleop stack

• Status: ACTIVE. Dated README news:
  • 2025-11-12: first release (Decoupled WBC for GR00T N1.5/N1.6)
  • 2026-02-19: GEAR-SONIC checkpoints
  • 2026-03-16: BONES-SEED dataset open-sourced
  • 2026-03-24: C++ stack update
  • 2026-04-10: SONIC training code
  • 2026-04-14: text-to-motion web demo
  • 2026-04-27: MotionBricks preview
  • 2026-05-07: end-to-end VLA workflow on G1
  • 2026-06-16: low-latency teleop checkpoint and Isaac Teleop/CloudXR docs
  • 2026-07-23: SONIC v1.1 checkpoint
• Who: NVIDIA GEAR lab. SONIC co-first authors are Zhengyi Luo (author of PHC), Ye Yuan, Tingwu Wang and Chenran Li. Project leads are Linxi "Jim" Fan and Yuke Zhu.
• Links: GitHub NVlabs/GR00T-WholeBodyControl; HF nvidia/GEAR-SONIC; arXiv 2511.07820.
• GitHub stats (fetched today): about 3k stars, 452 forks, no GitHub releases, 34 open issues.
• License: source code Apache-2.0; weights under the NVIDIA Open Model License. The HF card and a March 2026 news article both say commercial use is allowed, subject to NVIDIA's terms.
• What it is:
  • Decoupled WBC (used by GR00T N1.5/N1.6): an RL lower body plus an IK upper body.
  • SONIC is one universal tracking policy:
    • Trained with PPO.
    • Data: 700 h of in-house mocap, more than 100M frames at 50 Hz, 170 subjects of 145–199 cm.
    • Scaled from 1.2M to 42M parameters; 9k GPU-hours (128 GPUs for 3 days).
    • Uses a quantized (FSQ) motion-token interface.
    • A kinematic planner runs in under 5 ms on a laptop and 12 ms on a Jetson Orin, replanning every 100 ms.
    • Inputs: PICO whole-body VR, 3-point VR, video (pose estimation at 60 fps or more), text, music and VLA outputs.
  • Deployment:
    • C++/TensorRT stack with ONNX encoder, decoder and planner.
    • Three checkpoint variants: default (10 SMPL frames at 20 ms, about 200 ms lookahead), low-latency (4 frames, about 80 ms) and v1.1 (normalized to robot heading).
    • Robot: Unitree G1 (29-DoF action decoder). Issue #259 mentions H2 assets, so H2 support may be in progress (unverified).
    • Training and evaluation run in Isaac Lab; MuJoCo is used for sim2sim.
• Adoption / usage data:
  • HF GEAR-SONIC: 2,134 downloads in the last month.
  • EgoHumanoid (OpenDriveLab, RSS 2026) uses GR00T WBC as its low-level controller.
  • The issue tracker shows users porting it to new robots (#252 "Poor training on new robot") and running it in an outside disturbance benchmark (#269).
• Benchmarks (self-reported, arXiv 2511.07820):
  • Real world: 100% success on 50 varied trajectories (dance, jumps, loco-manipulation).
  • 3-point VR teleop over 300 trajectories: 121.9 ms mean latency; right-wrist position error 6 cm mean (13.3 cm p95); orientation error 0.145 rad (0.267 rad p95).
  • VLA fine-tune (apple-to-plate): 95% over 20 trials.
  • The paper says it beats Any2Track, BeyondMimic and GMT on unseen motions; I could not extract the numbers.
  • Performance improves steadily as data (0.4M → 7.4M → 100M frames), parameters and GPUs are scaled.
• Trade-offs:
  • Strengths: the only open generalist that combines weights, a planner, multi-modal teleop and a production C++ deploy stack.
  • Weaknesses:
    • G1-centric.
    • Full-scale retraining costs about 9k GPU-hours.
    • The 700-h training corpus is not public; the public BONES-SEED set is 288 h and gated.
  • Users report sim2sim falls and missing observations (#233, #270) and G1 "High Temperature" warnings (#247).
  • Choose BeyondMimic or mjlab to train your own high-fidelity per-skill trackers cheaply or on other robots. Choose holosoma for multi-robot (G1 plus T1) work.
• Sources: https://arxiv.org/html/2511.07820v1 · https://github.com/NVlabs/GR00T-WholeBodyControl/releases · https://github.com/NVlabs/GR00T-WholeBodyControl/blob/main/README.md · https://github.com/NVlabs/GR00T-WholeBodyControl/issues · https://huggingface.co/nvidia/GEAR-SONIC · https://rits.shanghai.nyu.edu/ai/nvidia-open-sources-sonic-a-foundation-model-for-humanoid-whole-body-control/

BeyondMimic (whole_body_tracking + motion_tracking_controller) — compact, high-fidelity per-motion tracking with real-world deployment

• Status: SLOWING (best estimate).
  • arXiv v1 2025-08-11, v4 2025-11-13.
  • One release tag, v1.0.0 ("Add observation history lengths to ONNX metadata"). The fetch tool dated it "July 24, 2024", which is before the project existed, so the year is unverified.
  • 25 open issues, numbered up to #70.
  • The same tracking task is also maintained in mjlab, which is ACTIVE.
• Who: UC Berkeley Hybrid Robotics (Koushil Sreenath) and Stanford (C. Karen Liu). Lead author Qiayuan Liao; also Takara E. Truong, Xiaoyu Huang, Yuman Gao and Guy Tevet.
• Links: HybridRobotics/whole_body_tracking; HybridRobotics/motion_tracking_controller; https://beyondmimic.github.io; arXiv 2508.08241.
• GitHub stats:
  • whole_body_tracking: 2.4k stars, 333 forks.
  • motion_tracking_controller: 540 stars, 59 forks, 51 commits.
• License: whole_body_tracking not confirmed (its main page could not be read). motion_tracking_controller is MIT.
• What it is:
  • Tracking objective: anchor-body pose error plus root velocity in the root frame.
  • Joint impedance tuned to a 10 Hz natural frequency; actions are normalized joint-position setpoints.
  • Trained in Isaac Lab. Highly dynamic skills (aerial cartwheels, spin-kicks, flip-kicks, sprinting) transfer zero-shot to a real G1.
  • The paper also describes a guided latent-diffusion layer (joystick, obstacle avoidance, inpainting). Only the tracking pipeline is open-sourced.
  • Data: LAFAN1. The project page shows 24 clips from 14 sequences of about 3 minutes.
  • Deployment: a ROS 2 Jazzy C++/ONNX controller on legged_control2, for MuJoCo sim and the real G1.
• Adoption:
  • Booster Robotics' booster_train "adapts the fabulous BeyondMimic" for the K1 robot.
  • robot_lab ships a BeyondMimic G1 tracking task.
  • mjlab has a motion-tracking task, co-authored by Liao.
  • SONIC uses it as a baseline.
• Benchmarks: I could not extract quantitative tables. The authors name a "planning-control gap" and say online replanning is "yet to be validated in the real world."
• Trade-offs:
  • Best fidelity for showcase motions. But it is one policy per motion, not a generalist (compare SONIC, GMT, TWIST).
  • Issues show Isaac Lab version breakage (#60, #70), difficulty porting to other robots (#61, #66 on per-joint natural frequency), and missing hand actuator configs (#63, #67).
• Sources: https://arxiv.org/abs/2508.08241 · https://arxiv.org/html/2508.08241v2 · https://beyondmimic.github.io/ · https://github.com/HybridRobotics/whole_body_tracking/releases · https://github.com/HybridRobotics/whole_body_tracking/issues · https://github.com/HybridRobotics/motion_tracking_controller

mjlab — Isaac Lab's manager-based API on MuJoCo Warp (GPU)

• Status: ACTIVE. PyPI 1.6.0 (2026-08-09), 1.5.x (June–July 2026), 1.0.0 (2026-02-02), first release 2025-09-26. 1,113 commits.
• Who: Kevin Zakka, Qiayuan Liao, Brent Yi, Louis Le Lay, Koushil Sreenath, Pieter Abbeel (UC Berkeley).
• Links: mujocolab/mjlab; https://pypi.org/project/mjlab/
• GitHub stats: about 3k stars, 523 forks.
• License: Apache-2.0.
• What it is:
  • Isaac Lab-style environment building blocks running directly on MuJoCo Warp, with direct access to MuJoCo data.
  • Built-in tasks: G1 velocity tracking, humanoid motion imitation and tracking, and sanity-check agents.
  • Training needs an NVIDIA GPU; macOS works for evaluation only. Multi-GPU and W&B are supported.
• Adoption:
  • BFM-Zero runs its distributed training in mjlab (about 6 h, per its README).
  • YAHMP (Inria, July 2026; arXiv 2607.19903) is built on mjlab 1.2.0.
  • The MuJoCo Warp README lists mjlab as its PyTorch/Isaac Lab-style integration.
  • A "twist2_mjlab" port appears in a June 2026 search-result title (not opened).
• Benchmarks: none in the README.
  • The Inria YAHMP study on mjlab (G1; 11,151 training and 1,024 test motions from AMASS and OMOMO) reports:
    • Removing observation history raises base errors by 38–50%.
    • Removing reference joint velocities degrades base errors by 7–14%.
    • Hand-force randomization lets the robot hold 4 kg with 6.6° of elbow sag, versus 15.5° without it.
    • Teacher-student training gives only about 6% better joint-position error.
• Trade-offs vs Isaac Lab:
  • mjlab: lighter install, MuJoCo-native contacts, the same simulator for training and sim2sim.
  • Isaac Lab: RTX rendering, a USD and asset ecosystem, and a much larger community (8.2k stars).
  • Isaac Lab 3.0's Newton backend uses MuJoCo Warp as its main solver, so the physics gap is narrowing.
• Sources: https://github.com/mujocolab/mjlab · https://pypi.org/project/mjlab/ · https://arxiv.org/html/2607.19903v1 · https://github.com/hucebot/yahmp · https://github.com/google-deepmind/mujoco_warp

holosoma (Amazon FAR) — full-stack, multi-robot, multi-simulator humanoid sim2real (locomotion and whole-body tracking), plus OmniRetarget

• Status: ACTIVE (probable).
  • Released alongside arXiv 2512.01996 (December 2025).
  • The OmniRetarget dataset was published 2025-09-30, and the project page shows an ICRA 2026 award designation.
  • 42 open issues about deployment and retargeting. No GitHub releases; commit dates not visible.
• Who: Amazon FAR (Frontier AI & Robotics).
  • 15-minute paper: Younggyo Seo, Carmelo Sferrazza, Juyue Chen, Guanya Shi, Rocky Duan, Pieter Abbeel.
  • OmniRetarget leads: Lujie Yang, Xiaoyu Huang, Zhen Wu, with MIT, Berkeley, Stanford and CMU.
• Links: amazon-far/holosoma; https://omniretarget.github.io; arXiv 2512.01996 and 2509.26633; HF omniretarget/OmniRetarget_Dataset.
• GitHub stats: 1.7k stars, 268 forks.
• License: Apache-2.0 (LICENSE file). The OmniRetarget dataset on HF is MIT.
• What it is:
  • Three packages:
    • holosoma: training for velocity locomotion and whole-body tracking
    • holosoma_inference: one pipeline shared by sim2sim and sim2real
    • holosoma_retargeting: converts mocap to robot motion "while preserving interactions with objects and terrain" (the OmniRetarget method)
  • Robots: G1 and T1.
  • Simulators: Isaac Gym, Isaac Sim (Ubuntu 22.04+) and MuJoCo Warp; MuJoCo for evaluation.
  • Algorithms: PPO and FastSAC (also FastTD3 in the paper).
  • Data: OMOMO and LAFAN.
• Adoption: 268 forks; the OmniRetarget dataset had 1,149 downloads in the last month. Issues come from users deploying on T1 and G1.
• Benchmarks (self-reported):
  • Sim-to-real locomotion trained in 15 minutes on one RTX 4090 (FastSAC/FastTD3), on G1 and T1.
  • Whole-body tracking on 4× L40s with 16,384 environments; a dance of more than 2 minutes on a real G1.
  • OmniRetarget versus PHC, GMR and VideoMimic on robot-object interaction:
    • penetration 0.00±0.01 s with max depth 1.34±0.34 cm
    • zero foot skating
    • downstream RL success 82.20%±9.74% (robot-object) and 94.73% (robot-terrain)
  • The paper generated more than 8 h of data (the project page says over 9 h). Only 4.0 h is public, and retargeted LAFAN1 is withheld "due to licensing restrictions."
• Trade-offs:
  • Strengths: the most complete open alternative to the NVIDIA stack for more than one robot and simulator; off-policy RL for fast iteration; the only open contact-preserving retargeter.
  • Weaknesses: no released generalist or loco-manipulation checkpoint (users ask for one); Isaac Gym legacy support.
  • Issues show T1 standing sway, G1 ankle limits that don't match the URDF, and requests for terrain-augmented motions (for example stairs from flat LAFAN clips), upper-body-only retargeting, and more mocap (a wall-flip request).
• Sources: https://github.com/amazon-far/holosoma/releases · https://github.com/amazon-far/holosoma/blob/main/LICENSE · https://github.com/amazon-far/holosoma/blob/main/README.md · https://github.com/amazon-far/holosoma/issues · https://arxiv.org/html/2512.01996v1 · https://arxiv.org/html/2509.26633v1 · https://omniretarget.github.io/ · https://huggingface.co/datasets/omniretarget/OmniRetarget_Dataset

GMR (General Motion Retargeting) — real-time CPU retargeting from human motion to many humanoids

• Status: SLOWING by maintainer news, but the community is active.
  • Last dated README news: 2026-01-21 (Xsens BVH) and 2026-01-12 (Fourier GR3, the 17th robot).
  • 73 open issues, including a proposal for BONES-SEED SOMA BVH input. BONES-SEED was released in March 2026, so these issues are recent.
• Who: Stanford (C. Karen Liu, Jiajun Wu). Yanjie Ze (now at Amazon FAR) maintains it; the tech report's first author is João Pedro Araújo, with Pei Xu.
• Links: YanjieZe/GMR; arXiv 2510.02252 ("Retargeting Matters") and 2505.02833.
• GitHub stats: 2.4k stars, 418 forks, no releases.
• License: MIT.
• What it is:
  • Method: non-uniform local scaling followed by a two-stage optimization. Root translation must be scaled uniformly to avoid foot sliding.
  • 17 humanoids, including G1/H1/H1-2, Booster T1/K1, HighTorque Hi, Galaxea R1 Pro, Kuavo, Berkeley Humanoid Lite, PND Adam Lite, Tienkung, PAL Talos and Fourier N1/GR3.
  • Inputs: SMPL-X (AMASS, OMOMO), BVH (LAFAN1, Nokov), OptiTrack FBX, Xsens BVH, and monocular video via GVHMR. Real-time PICO streaming via XRoboToolkit for TWIST2.
  • Kinematic only: no dynamics, and no object or terrain constraints.
• Adoption: the retargeter for TWIST and TWIST2 (7k of TWIST2's ~20k training clips); MimicKit ships converters from GMR's format; OmniRetarget uses it as a baseline.
• Benchmarks: on 21 LAFAN1 sequences (5 s to 2 min), tracking policies trained on GMR output succeeded 100% on Walk-2, Dance-1 and Dance-2. PHC's retargets got 23%, 0% and 0%. ProtoMotions' and Unitree's own retargets reached 99–100%.
• Trade-offs: the fastest way to cover many robots and to do real-time teleop. Use OmniRetarget when contacts or objects matter. Issues report arms placed too high (#157), G1 T-pose calibration with SMPL-X (#171), and requests for hands, upper-body-only, quadrupeds and Noitom data.
• Sources: https://github.com/YanjieZe/GMR/releases · https://github.com/YanjieZe/GMR/blob/master/README.md · https://github.com/YanjieZe/GMR/issues · https://arxiv.org/html/2510.02252v1

TWIST and TWIST2 — teleop to whole-body imitation, and portable data collection without mocap

• Status:
  • TWIST: borderline DORMANT. Fully open-sourced on 2025-09-29 (datasets plus teacher and student training); 7 commits.
  • TWIST2: SLOWING. Last README news 2025-12-02; 8 commits; the high-level policy "will be released in a separate repo."
• Who:
  • TWIST: Yanjie Ze, Zixuan Chen, João Pedro Araújo, Zi-ang Cao, Xue Bin Peng, Jiajun Wu, C. Karen Liu.
  • TWIST2: Ze (Amazon FAR), Siheng Zhao, Weizhuo Wang, Angjoo Kanazawa, Rocky Duan, Pieter Abbeel, Guanya Shi, Jiajun Wu, C. Karen Liu.
• Links: YanjieZe/TWIST; amazon-far/TWIST2; arXiv 2505.02833 and 2511.02832.
• GitHub stats: TWIST 818 stars, 76 forks; TWIST2 867 stars, 94 forks.
• License: both MIT. The TWIST2 datasets are CC BY 4.0 (per the paper).
• What it is:
  • TWIST: a teacher-student general motion tracker in Isaac Gym on G1. Originally driven by OptiTrack mocap; now uses GMR.
  • TWIST2:
    • PICO 4U headset plus ankle trackers and a custom 2-DoF neck ($250) for egocentric vision; the PICO setup costs about $1,000.
    • The tracker is trained on about 20k clips: 7k GMR-retargeted, 13k from TWIST, and 73 recorded with PICO.
    • A hierarchical visuomotor policy: Diffusion Policy on 224×224 RGB, with 64-step action chunks (about 2 s).
    • Robot: G1 29-DoF with Dex31 hands.
    • Two environments: Isaac Gym on Python 3.8, and GMR on Python 3.10+.
• Benchmarks (self-reported):
  • 100 demos in 15 minutes at almost 100% success.
  • Bimanual: 98 episodes in 18.5 minutes; mobile manipulation: 46 episodes in 19.5 minutes (both 100%).
  • Stated limits: struggles with sprinting; PICO body pose is less accurate than mocap at elbows and knees.
• Trade-offs: the cheapest complete open loop from teleop to data to policy for G1. SONIC also supports PICO teleop, with a stronger tracker but a heavier stack. HOMIE uses an exoskeleton and a non-commercial license.
• Sources: https://arxiv.org/html/2511.02832v1 · https://github.com/amazon-far/TWIST2 · https://github.com/YanjieZe/TWIST

ProtoMotions3 (NVIDIA) — GPU simulation and learning framework for simulated humans and humanoid robots

• Status: ACTIVE. It integrates Isaac Lab 3.0, Newton 1.0.0, BONES-SEED (March 2026) and Kimodo (April 2026). No GitHub releases.
• Who: NVIDIA Research: Chen Tessler, Yifeng Jiang, Xue Bin Peng, Erwin Coumans, Yi Shi, Haotian Zhang, Davis Rempe, Gal Chechik, Sanja Fidler.
• Links: NVlabs/ProtoMotions.
• GitHub stats: the main page showed 2.4k stars and 461 forks; the /releases page the same day showed 2.1k and 409. Treat it as roughly 2.1–2.4k.
• License: Apache-2.0.
• What it is:
  • Backends: Isaac Gym Preview 4, Isaac Lab 3.0, Newton 1.0.0, MuJoCo 3+, and Genesis (untested).
  • Characters and robots: SMPL humanoid, G1, H1_2.
  • Algorithms: MaskedMimic, ADD (adversarial differential discriminator), AMP, PPO.
  • Data: AMASS, BONES-SEED (~142K motions), PHUMA, Kimodo outputs.
  • Retargeting via PyRoki. A general tracker trained on BONES-SEED transfers zero-shot to a real G1.
• Benchmarks (self-reported): all of AMASS (40+ h) trained in 12 h on 4× A100. The BONES run used 24× A100 with 13K motions per GPU.
• Trade-offs: the widest simulator support and a strong character-animation and generative-control heritage, so it suits research on generalists. It is heavier than mjlab. For a ready G1 controller, SONIC's weights are more direct.
• Sources: https://github.com/NVlabs/ProtoMotions · https://github.com/NVlabs/ProtoMotions/releases · https://github.com/NVlabs/ProtoMotions/blob/main/README.md

MimicKit — Xue Bin Peng's reference suite of motion-imitation and motion-prior algorithms

• Status: SLOWING on verifiable evidence (arXiv v1 2025-10-15, v4 2026-01-18), but probably ACTIVE: it has Newton backend support and an SFU news item from March 2026 (search title only).
• Who: Xue Bin Peng (SFU / NVIDIA).
• Links: xbpeng/MimicKit; arXiv 2510.13794.
• GitHub stats: 2.3k stars, 294 forks, 194 commits.
• License: Apache-2.0.
• What it is: DeepMimic, AMP, AWR, ASE, LCP, ADD and SMP on Isaac Gym, Isaac Lab and Newton. Demos on humanoid characters, G1 and Go2. Converters from GMR and SMPL/AMASS formats.
• Trade-offs: clean, canonical implementations for algorithm research and teaching. It is not a hardware-deployment stack.
• Sources: https://github.com/xbpeng/MimicKit · https://arxiv.org/abs/2510.13794

rsl_rl (ETH RSL) — the default on-policy RL library for legged and humanoid robots

• Status: ACTIVE. 13 PyPI releases in 2026:
  • 5.5.1 (2026-09-09): multi-GPU fixes
  • 5.5.0 (2026-08-27): bf16 mixed precision for PPO and distillation
  • 5.4.0 (2026-05-27): pluggable loggers
  • 5.0.0 (2026-02-28)
• Who: ETH Zurich Robotic Systems Lab; maintainers Clemens Schwarke and Mayank Mittal.
• Links: leggedrobotics/rsl_rl; PyPI rsl-rl-lib.
• GitHub stats: 3k stars, 685 forks.
• License: BSD-3-Clause.
• What it is: PPO plus student-teacher distillation, with multi-GPU support.
• Adoption: dependency of unitree_rl_gym, HumanoidVerse, human2humanoid, HumanPlus, robot_lab (symmetry augmentation) and WBC-AGILE (compatibility patch).
• Trade-offs: simple and fast. For off-policy sample efficiency use FastSAC/FastTD3 (holosoma); for JAX use Brax PPO (Playground) or LocoMuJoCo.
• Sources: https://pypi.org/project/rsl-rl-lib/ · https://github.com/leggedrobotics/rsl_rl/releases

legged_gym — the 2021 Isaac Gym template that most humanoid repos were forked from

• Status: DORMANT. The README deprecation note (dated 09.01.2024) says the environments moved to Isaac Lab and the repo gets "limited updates." 18 commits.
• Who: ETH RSL (Nikita Rudin et al.).
• Links: leggedrobotics/legged_gym; arXiv 2109.11978.
• GitHub stats: 3.1k stars, 586 forks.
• License: BSD-3-Clause (not re-verified today).
• What it is: Isaac Gym Preview 3 environments. It is the template for Humanoid-Gym, unitree_rl_gym, HumanPlus and H2O.
• Known issue (README): GPU contact forces are unreliable on triangle-mesh terrain.
• Trade-offs: use only to reproduce older work.
• Sources: https://github.com/leggedrobotics/legged_gym

unitree_rl_gym and unitree_rl_lab (Unitree) — vendor RL locomotion stacks

• Status:
  • unitree_rl_gym: not verifiable (no releases; 42 commits).
  • unitree_rl_lab: probably ACTIVE or SLOWING. It needs Isaac Lab ≥2.3.0 and Isaac Sim ≥5.1.0, which are late-2025 or later; 77 commits.
• Who: Unitree Robotics.
• Links: unitreerobotics/unitree_rl_gym; unitreerobotics/unitree_rl_lab.
• GitHub stats: unitree_rl_gym 3.5k stars, 579 forks; unitree_rl_lab 1.3k stars, 317 forks.
• License: unitree_rl_gym BSD-3-Clause; unitree_rl_lab Apache-2.0.
• What it is:
  • rl_gym: Go2, G1, H1 and H1_2 on Isaac Gym, then MuJoCo sim2sim, then real robot via unitree_sdk2_python, plus a C++ LibTorch example for G1.
  • rl_lab: Isaac Lab tasks for Go2, H1 and G1-29dof; sim2sim via unitree_mujoco; a C++ g1_ctrl for the real robot.
  • Both READMEs I fetched are about locomotion; neither mentions general motion tracking.
• Trade-offs: the standard starting point for Unitree locomotion. For tracking, people move to BeyondMimic, mjlab, SONIC or holosoma.
• Sources: https://github.com/unitreerobotics/unitree_rl_gym · https://github.com/unitreerobotics/unitree_rl_lab

Humanoid-Gym (RobotEra) — Isaac Gym humanoid locomotion with MuJoCo sim2sim

• Status: DORMANT (37 commits; 2024-era).
• Who: RobotEra (Xinyang Gu, Yen-Jen Wang, Jianyu Chen).
• Links: roboterax/humanoid-gym; arXiv 2404.05695.
• GitHub stats: 2.1k stars, 251 forks.
• License: not confirmed; no LICENSE file at the expected path.
• What it is: zero-shot sim2real on XBot-S (1.2 m) and XBot-L (1.65 m). It popularized the habit of validating in MuJoCo before going to hardware.
• Sources: https://github.com/roboterax/humanoid-gym

booster_gym and booster_train (Booster Robotics) — vendor stacks for T1 and K1

• Status:
  • booster_gym: legacy (7 commits). Its README now points to a new pipeline made of booster_train (Isaac Lab), booster_deploy and booster_assets.
  • booster_train: status unverified (6 commits).
• Links: BoosterRobotics/booster_gym (arXiv 2506.15132, from search title); BoosterRobotics/booster_train.
• GitHub stats: booster_gym 304 stars, 54 forks; booster_train 56 stars, 17 forks.
• License: booster_gym Apache-2.0; booster_train not shown.
• What it is:
  • booster_gym: Isaac Gym plus MuJoCo; T1, and newly K1.
  • booster_train: BeyondMimic tracking adapted for K1 on Isaac Sim 5.0 / Isaac Lab 2.2, exporting TorchScript or ONNX.
• Sources: https://github.com/BoosterRobotics/booster_gym · https://github.com/BoosterRobotics/booster_gym/blob/main/LICENSE · https://github.com/BoosterRobotics/booster_train

HumanoidVerse and ASAP (CMU LeCAR Lab) — multi-simulator framework, and learned sim-to-real dynamics alignment

• Status:
  • HumanoidVerse: DORMANT. Last dated news 2025-02-04; 9 commits.
  • ASAP: DORMANT (probable). RSS 2025; no 2026 evidence.
• Who: LeCAR Lab (PI Guanya Shi).
  • HumanoidVerse is led by Jiawei Gao, with Tairan He, Wenli Xiao, Yuanhang Zhang and others.
  • ASAP's lead authors are Tairan He, Jiawei Gao, Wenli Xiao and Yuanhang Zhang, together with NVIDIA.
• Links: LeCAR-Lab/HumanoidVerse; LeCAR-Lab/ASAP; arXiv 2502.01143.
• GitHub stats: HumanoidVerse 470 stars, 32 forks (low; reported as fetched); ASAP 2.1k stars, 197 forks.
• License: both MIT.
• What it is:
  • HumanoidVerse: one codebase for Isaac Gym, Isaac Sim 4.2 and Genesis 0.2.1; H1 and G1; locomotion. Its README lists tracking as "planned."
  • ASAP, built on it:
    • phase-based motion tracking plus a delta-action model, which learns residual actions so simulation matches real dynamics
    • retargeting; MuJoCo sim2sim; real-robot deployment with UnitreeSDK
    • G1 29-DoF; data from AMASS plus video via TRAM
• Adoption: KungfuBot (PBHC) is built on HumanoidVerse.
• Trade-offs: ASAP is the main open implementation of learned real-to-sim correction. It is stuck on 2025-era simulators.
• Sources: https://github.com/LeCAR-Lab/HumanoidVerse · https://github.com/LeCAR-Lab/ASAP

HOVER (NVIDIA GEAR + CMU) — multi-mode command space distilled into one WBC policy

• Status: DORMANT. 15 commits; tested only with Isaac Lab 2.0.0. NVIDIA's later WBC work is GR00T WBC / SONIC.
• Links: NVlabs/HOVER; arXiv 2410.21229.
• GitHub stats: 758 stars, 90 forks.
• License: Apache-2.0.
• What it is: teacher-student distillation on H1, using AMASS retargeted to H1. Includes a MuJoCo sim2sim wrapper and sim-to-real tools.
• Sources: https://github.com/NVlabs/HOVER

H2O / OmniH2O (human2humanoid) — real-time teleoperation of the whole humanoid body

• Status: DORMANT (IROS and CoRL 2024; 30 commits).
• Who: Tairan He and Zhengyi Luo (CMU).
• Links: LeCAR-Lab/human2humanoid; arXiv 2406.08858.
• GitHub stats: 1.1k stars, 100 forks.
• License: CC BY-NC 4.0 (non-commercial).
• What it is: Isaac Gym; H1; AMASS retargeting; teacher-student training; real-robot deployment.
• Sources: https://github.com/LeCAR-Lab/human2humanoid

PHC (Perpetual Humanoid Control) — physics-based SMPL tracking that recovers from failures; the main AMASS-scale baseline

• Status: DORMANT. Last dated update in August 2025 (Isaac Lab evaluation code). Earlier: offline dataset generation (December 2024) and H1/G1 support (October 2024).
• Who: Zhengyi Luo (CMU, now NVIDIA). ICCV 2023.
• Links: ZhengyiLuo/PHC.
• GitHub stats: 1.3k stars, 125 forks, 200 commits.
• License: BSD 3-Clause Clear. The SMPL body models it uses carry their own non-commercial terms.
• What it is: Isaac Gym; SMPL/SMPL-X; 11,313 cleaned AMASS sequences; scripts for retargeting to your own humanoid.
• Trade-offs: its retargeter did very badly in GMR's study (0% success on dance clips).
• Sources: https://github.com/ZhengyiLuo/PHC · https://github.com/ZhengyiLuo/PHC/blob/master/LICENSE

GMT (General Motion Tracking) — a single G1 tracker for diverse motions

• Status: DORMANT. 4 commits. Only a pretrained checkpoint plus MuJoCo sim2sim is released; the README still says data processing and retargeter code "will be released soon."
• Who: UCSD (Xiaolong Wang) and SFU (Xue Bin Peng); lead Zixuan Chen.
• Links: zixuan417/humanoid-general-motion-tracking; arXiv 2506.14770.
• GitHub stats: 439 stars, 20 forks.
• License: Apache-2.0.
• Trade-offs: commonly used as a baseline (SONIC, BeyondMimic), but you cannot retrain it.
• Sources: https://github.com/zixuan417/humanoid-general-motion-tracking · LICENSE at /blob/master/LICENSE

BFM-Zero (CMU LeCAR + Meta) — promptable behavioral foundation model trained with unsupervised RL

• Status: ACTIVE (probable). ICLR 2026; 104 commits; released in stages. Full training and evaluation pipelines are out; a minimal RTX 4090 training script is still pending.
• Links: LeCAR-Lab/BFM-Zero; HF LeCAR-Lab/BFM-Zero; arXiv 2511.04131.
• GitHub stats: 715 stars, 67 forks.
• License: CC BY-NC 4.0.
• What it is: a forward-backward representation policy that can be prompted for tracking, goal-reaching or reward optimization, exported to ONNX. Runs on Isaac Sim or MuJoCo; distributed training in mjlab (about 6 h); teleop via UFO. Robot: G1 (per the paper).
• Trade-offs: the most interesting "prompt instead of retrain" controller, but the non-commercial license blocks product use.
• Sources: https://github.com/LeCAR-Lab/BFM-Zero

HOMIE (Shanghai AI Lab + CUHK) — exoskeleton-cockpit teleop with RL loco-manipulation

• Status: DORMANT. Everything released in February 2025; 16 commits.
• Links: InternRobotics/OpenHomie; arXiv 2502.13013.
• GitHub stats: 619 stars, 49 forks.
• License: CC BY-NC-SA 4.0.
• What it is: G1 with Dex3 hands; Isaac Gym Preview 4; hardware designs for isomorphic exoskeleton arms and sensing gloves (about $0.5k); RealSense cameras.
• Sources: https://github.com/InternRobotics/OpenHomie

AMO (UCSD) — adaptive motion optimization for hyper-dexterous whole-body control

• Status: DORMANT. RSS 2025; 4 commits; deployment files and pretrained weights only, no training code.
• Links: OpenTeleVision/AMO; arXiv 2505.03738.
• GitHub stats: 379 stars, 22 forks.
• License: Apache-2.0.
• What it is: G1, visualized in MuJoCo.
• Sources: https://github.com/OpenTeleVision/AMO · /blob/main/LICENSE

FALCON (CMU LeCAR + Field AI) — force-adaptive humanoid loco-manipulation

• Status: SLOWING. Its news item is acceptance as an oral at L4DC 2026; 26 commits.
• Links: LeCAR-Lab/FALCON; arXiv 2505.06776.
• GitHub stats: 435 stars, 49 forks.
• License: MIT.
• What it is: Isaac Gym Preview 4; G1 29-DoF and Booster T1 29-DoF; real-robot code for both vendors' SDKs. One of the few releases that runs on two robot brands.
• Sources: https://github.com/LeCAR-Lab/FALCON

KungfuBot / PBHC (TeleAI and others) — tracking highly dynamic motions (kung fu, dance) taken from video

• Status: SLOWING. Code June 2025, NeurIPS 2025, general-tracking support October 2025; 40 commits. KungfuBot2 is a follow-up.
• Who: Weiji Xie, Jinrui Han, Chenjia Bai (China Telecom TeleAI), with SJTU and others.
• Links: TeleHuman/PBHC; arXiv 2506.12851.
• GitHub stats: 1.1k stars, 142 forks.
• License: CC BY-NC 4.0.
• What it is: video to SMPL via GVHMR (also LAFAN and AMASS), physics-based filtering, retargeting, then RL tracking in Isaac Gym on HumanoidVerse. Robot: G1.
• Sources: https://github.com/TeleHuman/PBHC

VideoMimic (UC Berkeley) — monocular video to reconstructed human and scene in sim, then a context-aware humanoid policy

• Status: SLOWING. Real-to-sim pipeline 2025-07-06; simulation and preliminary sim2real code 2025-09-15; README reports a Best Student Paper award on 2025-09-30.
• Who: Arthur Allshire, Hongsuk Choi, Junyi Zhang, David McAllister, Anthony Zhang, Chung Min Kim, Trevor Darrell, Pieter Abbeel, Jitendra Malik, Angjoo Kanazawa.
• Links: hongsukchoi/VideoMimic; arXiv 2505.03729.
• GitHub stats: 831 stars, 66 forks.
• License: MIT.
• What it is: a four-stage training pipeline (mocap pretraining, scene tracking, distillation, RL fine-tuning) with a C++ deploy path; the video dataset is on Google Drive. Robot: G1 (per the paper).
• Sources: https://github.com/hongsukchoi/VideoMimic

HDMI (CMU LeCAR) — interactive whole-body skills learned from monocular RGB human videos

• Status: SLOWING (unverified; Isaac Lab 2.2-era).
• Links: LeCAR-Lab/HDMI; arXiv 2509.16757.
• GitHub stats: 664 stars, 49 forks, 18 commits.
• License: not shown.
• What it is: Isaac Sim 4.5 / Isaac Lab 2.2; G1; teacher and student training and evaluation code.
• Sources: https://github.com/LeCAR-Lab/HDMI

HumanPlus (Stanford) — humanoid shadowing transformer plus imitation transformer

• Status: DORMANT (CoRL 2024; 1 commit).
• Links: MarkFzp/humanplus.
• GitHub stats: 849 stars, 122 forks.
• License: not specified.
• What it is: HST (RL in Isaac Gym on legged_gym and rsl_rl) plus HIT; H1 with Inspire hands.
• Sources: https://github.com/MarkFzp/humanplus

ExBody2 (UCSD) — expressive whole-body tracking

• Status: paper only (arXiv 2412.13196). I found no official code repository (search showed only paper and news pages). Treat it as closed.

Humanoid-GPT (Galbot) — GPT-style zero-shot motion tracker (CVPR 2026)

• Status: ACTIVE (new; arXiv 2606.03985, June 2026; 10 commits).
• Who: Galbot / GalaxyGeneralRobotics; lead Zekun Qi.
• Links: GalaxyGeneralRobotics/Humanoid-GPT.
• GitHub stats: 457 stars, 34 forks.
• License: Apache-2.0.
• What it is: a causal transformer with RoPE, trained on 2 billion motion frames (public mocap plus large in-house recordings). G1 29-DoF; MuJoCo/MJX. Inference code and checkpoints are released; training code and data are listed as TODO.
• Sources: https://github.com/GalaxyGeneralRobotics/Humanoid-GPT

WBC-AGILE (NVIDIA Isaac) — production-style Isaac Lab workflow for humanoid loco-manipulation

• Status: ACTIVE. Built on Isaac Lab v3.0.0-beta2 (June 2026+); arXiv 2603.20147 (March 2026); 53 commits.
• Who: Huihua Zhao, Rafael Cathomen, Lionel Gulich, Efe Arda Ongan, Michael Lin, Shalin Jain, Wei Liu and others.
• Links: nvidia-isaac/WBC-AGILE.
• GitHub stats: 346 stars, 42 forks.
• License: Apache-2.0 (its RSL-RL compatibility patch is BSD-3).
• What it is:
  • Robots: G1 and T1, with sim-and-real videos.
  • Tasks: velocity, height control, sit and stand, teleop, dancing.
  • An evaluation framework (random rollouts, deterministic scenarios, motion metrics, HTML reports), sim-to-MuJoCo checks, and OSMO distributed training.
• Trade-offs: the most "engineering-grade" Isaac Lab 3 template. It does not ship a generalist tracker.
• Sources: https://github.com/nvidia-isaac/WBC-AGILE

robot_lab — community Isaac Lab extension covering many robots

• Status: SLOWING on releases: v2.3.2 on Feb 3 (shown as 2025 by the fetch tool; it follows v2.3.0, which targets Isaac Sim 5.1, so the date is most likely 2026-02-03). 196 commits.
• Who: Ziqi Fan (fan-ziqi).
• GitHub stats: 1.9k stars, 236 forks.
• License: Apache-2.0.
• What it is:
  • Humanoids: G1/H1, Fourier GR1T1/T2, Booster T1, RobotEra Xbot, OpenLoong, RoboParty ATOM01, MagicBot.
  • Tasks: velocity, BeyondMimic tracking (G1), experimental AMP dance, handstand, distillation with symmetry.
• Trade-offs: broadest list of robots in Isaac Lab. It is a hobbyist-to-research bridge rather than a lab release.
• Sources: https://github.com/fan-ziqi/robot_lab · /releases

LocoMuJoCo (TU Darmstadt) — imitation-learning benchmark for whole-body locomotion

• Status: SLOWING. PyPI 1.0.0 was 2025-04-17. GitHub v1.1.0 adds a MuJoCo Warp backend and needs MuJoCo 3.5+; it is dated Mar 10 (shown as 2025, most likely 2026). 723 commits.
• Who: Firas Al-Hafez, Davide Tateo, Jan Peters.
• GitHub stats: 1.5k stars, 154 forks.
• License: MIT.
• What it is: 12 humanoid and 4 quadruped environments, including 4 biomechanical human models; more than 22,000 mocap samples per humanoid; LAFAN1 and AMASS loaders; single-file JAX PPO, GAIL, AMP and DeepMimic.
• Sources: https://github.com/robfiras/loco-mujoco · /releases · https://pypi.org/project/loco-mujoco/

HumanoidBench (UC Berkeley) — simulated humanoid locomotion and manipulation benchmark

• Status: DORMANT (RSS 2024; 45 commits).
• GitHub stats: 799 stars, 130 forks.
• License: MIT (composite).
• What it is: H1 with Shadow hands and G1 with three-finger hands; 27 tasks; DreamerV3, TD-MPC2, SAC and PPO baselines; MuJoCo; no sim2real.
• Sources: https://github.com/carlosferrazza/humanoid-bench · /blob/main/LICENSE

MuJoCo Playground (Google DeepMind + collaborators) — GPU environments on MJX and MuJoCo Warp

• Status: borderline ACTIVE/SLOWING. PyPI 0.2.0 (2026-03-16) and 0.1.0 (2026-01-08); 221 commits.
• Who: Kevin Zakka, Baruch Tabanpour, Qiayuan Liao, …, Koushil Sreenath, Carmelo Sferrazza, Yuval Tassa, Pieter Abbeel.
• GitHub stats: 2.2k stars, 355 forks.
• License: Apache-2.0.
• What it is: the abstract claims policies train "in minutes on a single GPU" with zero-shot sim2real from state and pixel inputs. Training uses JAX PPO and rsl-rl PPO. I did not re-verify which humanoid models are included.
• Sources: https://github.com/google-deepmind/mujoco_playground · https://pypi.org/project/playground/ · https://arxiv.org/abs/2502.08844

K-Scale ksim and kinfer — JAX/MuJoCo RL and inference for K-Bot

• Status: DISCONTINUED. K-Scale Labs shut down in early November 2025 and open-sourced its IP. ksim's last release was 0.2.10 (2025-08-18); the repo is not archived. CEO Benjamin Bolte later joined OpenAI (humanoidsdaily). I did not fetch kinfer.
• GitHub stats: ksim 225 stars, 39 forks, MIT.
• Sources: https://github.com/kscalelabs/ksim · https://www.humanoidsdaily.com/news/watch-k-scale-labs-ceo-explains-shutdown-i-bet-the-farm-on-the-k-bot-launch

B. Model-based and classical control

Pinocchio — rigid-body dynamics with analytical derivatives; the base layer of the European model-based stack

• Status: ACTIVE. PyPI 4.1.0 on 2026-07-08 (URDF v1.2). v4.0.0 (April 13; most likely 2026) added loop-constraint ABA, a new constraint API, and ADMM/PGS solvers.
• Who: Inria Willow and LAAS-CNRS (Justin Carpentier, Nicolas Mansard).
• GitHub stats: 3.8k stars, 578 forks.
• License: BSD-2-Clause.
• What it is: used by Crocoddyl, TSID, Aligator, Pink, PlaCo and OCS2's URDF helpers.
• Sources: https://github.com/stack-of-tasks/pinocchio/releases · https://pypi.org/project/pin/ · /blob/devel/LICENSE

Crocoddyl — DDP/FDDP optimal control for multi-contact locomotion

• Status: ACTIVE. v3.2.1 (2026-05-10) fixed Pinocchio v4 compatibility; v3.2.0 (2025-12-09); v3.1.0 (2025-10-03).
• Who: LAAS-CNRS / Carlos Mastalli et al.
• GitHub stats: 1.3k stars, 215 forks.
• License: BSD-3-Clause.
• Sources: https://github.com/loco-3d/crocoddyl/releases · /blob/devel/LICENSE

Aligator — constrained trajectory optimization (proximal augmented-Lagrangian DDP)

• Status: ACTIVE (probable). v0.19.0 adds tentative Pinocchio 4.0 support; dated Apr 17 (the tool showed 2025, but Pinocchio 4 shipped in April 2026).
• Who: Inria (Simple-Robotics).
• GitHub stats: 317 stars, 47 forks.
• License: BSD-2-Clause.
• Sources: https://github.com/Simple-Robotics/aligator/releases · /blob/main/LICENSE

TSID — task-space inverse dynamics QP

• Status: ACTIVE. v1.10.0 (2026-04-14); v1.9.0 (2025-09-29).
• Who: Andrea Del Prete et al.
• GitHub stats: 349 stars, 96 forks.
• License: BSD-2-Clause.
• Sources: https://github.com/stack-of-tasks/tsid/releases · /blob/devel/LICENSE

OCS2 (ETH RSL) — switched-system MPC (SLQ, DDP, iLQR, SQP, IPM)

• Status: not verifiable. 7,471 commits; the main branch is ROS 1 and a ros2 branch exists.
• Who: ETH RSL (Farbod Farshidian et al.).
• GitHub stats: 1.5k stars, 329 forks.
• License: BSD-3-Clause.
• Sources: https://github.com/leggedrobotics/ocs2

Drake (Toyota Research Institute) — multibody dynamics, optimization and controller design

• Status: ACTIVE. v1.57.0 (2026-09-10), v1.56.0 (2026-08-14), v1.55.0 (2026-07-15).
• GitHub stats: 4.2k stars, 1.4k forks.
• License: BSD-3-Clause (not re-verified today).
• What it is: strong for MPC and trajectory optimization research and verification. Rarely used in the humanoid RL stacks above.
• Sources: https://github.com/RobotLocomotion/drake/releases

IHMC Open Robotics Software — whole-body momentum control and footstep planning (Valkyrie, Nadia, Alex)

• Status: not verifiable (65,271 commits; the visible release tag is 0.14.0-240126).
• GitHub stats: 325 stars, 110 forks.
• License: Apache-2.0.
• Trade-offs: a mature Java stack; little adoption outside IHMC.
• Sources: https://github.com/ihmcrobotics/ihmc-open-robotics-software

MuJoCo MPC (MJPC, Google DeepMind) — real-time predictive control (iLQG, gradient descent, predictive sampling), including a humanoid mocap-tracking task

• Status: not verifiable (1,562 commits; 39 open issues and 20 open PRs).
• GitHub stats: 1.7k stars, 282 forks.
• License: Apache-2.0.
• Sources: https://github.com/google-deepmind/mujoco_mpc

mink — differential IK on MuJoCo

• Status: SLOWING by releases (v1.1.0 in February 2026); 301 commits.
• Who: Kevin Zakka.
• GitHub stats: 1.5k stars, 115 forks.
• License: Apache-2.0.
• What it is: examples include G1, H1 and Apptronik Apollo; collision avoidance between any pair of geoms; inspired by Pink. Its role inside GMR is not verified here.
• Sources: https://github.com/kevinzakka/mink

Pink — weighted-QP differential IK on Pinocchio

• Status: ACTIVE. 4.4.0 (2026-09-09), 4.3.0 (2026-07-15), 4.2.0 (2026-04-20).
• Who: Stéphane Caron.
• GitHub stats: 851 stars, 70 forks.
• License: Apache-2.0.
• Sources: https://github.com/stephane-caron/pink · https://pypi.org/project/pin-pink/

PlaCo (Rhoban) — Pinocchio-based IK/ID QP with humanoid walking utilities

• Status: ACTIVE (PyPI 0.10.1 on 2026-09-28).
• GitHub stats: 353 stars, 29 forks, 1,330 commits.
• License: MIT.
• Sources: https://github.com/Rhoban/placo · https://pypi.org/project/placo/

legged_control — NMPC + WBC + state estimation on OCS2 and ros-control

• Status: DISCONTINUED. The README says "This software is not supported anymore! The authors are developing a completely new framework."
• Who: Qiayuan Liao. BeyondMimic's deploy stack depends on "legged_control2."
• GitHub stats: 1.8k stars, 366 forks.
• License: BSD-3-Clause.
• Sources: https://github.com/qiayuanl/legged_control

OpenLoong-Dyn-Control — MPC + WBC for the Qinglong (OpenLoong) humanoid

• Status: DORMANT (last version entry 2024.09.11).
• Who: Humanoid Robots (Shanghai) Co.
• GitHub stats: 347 stars, 70 forks.
• License: Apache-2.0.
• Sources: https://github.com/loongOpen/OpenLoong-Dyn-Control

ros2_control — the ROS 2 hardware and controller framework used by real-robot deploy stacks

• Status: not verified. No GitHub releases; it ships through ROS distributions.
• GitHub stats: 973 stars, 466 forks.
• License: Apache-2.0 (not re-verified).
• What it is: BeyondMimic's motion_tracking_controller is a ROS 2 Jazzy controller. I found no dedicated open humanoid WBC controller in ros2_controllers.
• Sources: https://github.com/ros-controls/ros2_control/releases

C. Simulation backends these stacks depend on (brief)

Sources: https://github.com/isaac-sim/IsaacLab/releases · https://github.com/newton-physics/newton · https://pypi.org/project/mujoco-warp/ · https://github.com/google-deepmind/mujoco_warp · https://pypi.org/project/genesis-world/ · https://github.com/Genesis-Embodied-AI/Genesis

D. Motion data these pipelines use, and its licenses

Key quantitative evidence that data quality matters:

• PHUMA: G1 tracking success was 92.7% when trained on PHUMA, versus 76.2% on AMASS, 46.1% on LAFAN1 and 50.6% on Humanoid-X. On unseen video motions: 82.9% versus 70.2% for AMASS.
• GMR: PHC-retargeted dance clips reached 0% tracking success.
• SONIC: performance keeps improving as data grows from 0.4M to 100M frames.

Sources: https://amass.is.tue.mpg.de/license.html · https://github.com/ubisoft/ubisoft-laforge-animation-dataset · https://huggingface.co/datasets/lvhaidong/LAFAN1_Retargeting_Dataset · https://github.com/lijiaman/omomo_release · https://www.roboticstomorrow.com/news/2026/03/17/bones-studio-to-release-bones-seed-the-first-multimodal-motion-dataset-purpose-built-for-humanoid-robotics/26276/ · https://huggingface.co/datasets/bones-studio/seed · https://huggingface.co/datasets/DAVIAN-Robotics/PHUMA · https://arxiv.org/html/2510.26236v1 · https://huggingface.co/datasets/omniretarget/OmniRetarget_Dataset · https://huggingface.co/nvidia/Kimodo-SOMA-SEED-v1.1 · https://huggingface.co/papers/2608.13555 · https://arxiv.org/html/2603.09170v1

Cross-cutting trade-offs

• Isaac Lab vs MuJoCo-based stacks (mjlab, Playground, LocoMuJoCo, Humanoid-GPT's MJX):
  • Isaac Lab: the largest community, NVIDIA assets and teleop (CloudXR), RTX rendering. But version churn between 2.x and 3.0 breaks downstream repos (BeyondMimic issues #60 and #70).
  • MuJoCo Warp stacks: lighter, deterministic MuJoCo contacts, and training and sim2sim in the same engine.
  • Isaac Lab 3.0 with Newton narrows the physics gap.
  • In practice most teams train in one and validate in MuJoCo; nearly every repo ships a MuJoCo sim2sim step.
• Per-motion tracking vs generalist tracking:
  • Per-motion (BeyondMimic, booster_train): highest fidelity, cheap per clip, but no generalization.
  • Generalists (SONIC, GMT, TWIST, Humanoid-GPT, ProtoMotions): need large, clean, retargeted data and much more compute (SONIC: 9k GPU-hours; ProtoMotions: 4× A100 for 12 h on AMASS).
• Motion tracking vs teleop-based data: trackers are the low-level substrate for teleop. TWIST2, SONIC-PICO and EgoHumanoid layer VR or egocentric capture on top of a tracker and then train VLA or Diffusion policies. The quality of the tracker caps the quality of the teleop data.
• RL vs MPC:
  • RL dominates dynamic whole-body behaviour on G1 and T1.
  • Model-based libraries survive inside retargeting and IK (Pink, mink, PlaCo, OmniRetarget's constrained optimization) and in hybrids. HybridMimic (arXiv 2603.06775, Booster T1) has an RL policy modulate a centroidal controller and reports 13% lower base-position tracking error than an RL baseline.
  • Integrated open MPC+WBC humanoid stacks are dormant (legged_control, OpenLoong).

Also notable (brief)

• YAHMP (hucebot/yahmp, Inria; 12 stars; Apache-2.0; July 2026; mjlab-based G1 tracker and ablation study) — ACTIVE. https://github.com/hucebot/yahmp
• EgoHumanoid (OpenDriveLab/EgoHumanoid, RSS 2026; 218 stars; Apache-2.0) — whole-body loco-manipulation learned from egocentric human demos; reports +51% over robot-only data in unseen environments (self-reported); low-level control is GR00T WBC — ACTIVE.
• ZeroWBC (arXiv 2603.09170; USTC, Shanghai AI Lab and others) — egocentric human data (Nymeria 300 h, HumanML3D, about 5 h self-collected with chest GoPro and mocap suit) generating motions for a G1 tracker; code status unclear.
• ReactiveBFM (zjwzcx/ReactiveBFM, CoRL 2026; 52 stars; CC BY-NC-SA 4.0; CUHK + Shanghai AI Lab) — diffusion motion planner on top of a tracker; data and evaluation code promised for September 2026 — ACTIVE.
• HumanTracker (GalaxyGeneralRobotics/HumanTracker; 32 stars per the HF page; Aug 2026) — benchmark plus learned human-preference metric — ACTIVE.
• MotionBricks (NVIDIA DAIR, SIGGRAPH 2026) — real-time latent generative motion; preview in GR00T WBC on 2026-04-27.
• Seen as search-result titles only (not opened; status unverified):
  • CLOT (arXiv 2602.15060, closed-loop global tracking for teleop)
  • SceneBot (arXiv 2606.27581, contact-prompted scene-interaction tracking)
  • "Make Tracking Easy: Neural Motion Retargeting" (arXiv 2603.22201)
  • HiPHI (arXiv 2608.16222, human motion and object-interaction benchmark)
  • amazon-far/PRISM (CoRL 2026, counterfactual video generation for loco-manipulation)
  • OpenDriveLab/WholebodyVLA (ICLR 2026)
  • LeCAR-Lab/SoFTA
  • twist2_mjlab (community port)
  • GuilhermeAsura/humanoid_repos_eval (a repo for running and comparing humanoid locomotion repos)

Comparison table

Gaps and pain points (with evidence), focused on a data, simulation or environment supplier

1. No commercially usable, robot-ready motion data at scale.
   • AMASS bans commercial training and derivative redistribution. LAFAN1 is NC-ND, and Unitree's LAFAN1 retarget inherits that.
   • Releases are being cut down because of licensing: OmniRetarget withheld retargeted LAFAN1, and PHUMA's HF release left out LAFAN1 and LocoMuJoCo motions.
   • The best-performing model was trained on 700 h of private mocap (SONIC); Galbot's Humanoid-GPT used 2B in-house frames.
   • BONES-SEED (288 h) is the largest public, robot-formatted set, but it is gated under a custom license with a commercial-licensing contact.
   • Opportunity: commercially licensed mocap delivered already retargeted and physics-validated per robot. PHUMA's 92.7% versus AMASS's 76.2% G1 tracking success shows that curation is worth paying for.
2. Retargeting is the bottleneck between data and policy.
   • GMR's study: tracking success swings from 0% to 100% depending only on the retargeter.
   • Users report:
     • arm offsets and T-pose calibration problems (GMR #157, #171)
     • requests for upper-body-only retargeting (GMR #176; holosoma)
     • hand and dexterous retargeting gaps (GMR #175; BeyondMimic #63/#67)
     • new mocap formats (Noitom, SOMA BVH)
     • terrain augmentation of flat clips (holosoma issues)
   • Opportunity: retargeting as a service, plus per-robot QA metrics.
3. Interaction data (objects, terrain, scenes) is scarce.
   • Public OmniRetarget data is 4.0 h; OMOMO depends on SMPL licensing; users ask for pretrained object-interaction policies and specific stunt mocap (holosoma issues).
   • 2026 work keeps moving to scene and contact tracking (SceneBot, HDMI, VideoMimic, HiPHI).
4. Egocentric plus body-pose capture is fragmented and inaccurate.
   • Rigs in use: TWIST2 (PICO 4U plus ankle trackers, about $1.25k), EgoHumanoid (PICO plus 5 trackers plus ZED Mini), ZeroWBC (chest GoPro plus mocap suit), HOMIE (about $0.5k exoskeleton and gloves).
   • TWIST2 reports PICO body pose is weak at elbows and knees and fails on sprinting.
   • EgoHumanoid reports +51% from human egocentric data, which suggests demand.
   • Opportunity: a standardized, calibrated rig and dataset pairing egocentric video, full-body pose and hands (a data-glove angle).
5. One robot dominates, and porting is painful.
   • Almost everything is G1-first. Users report "Poor training on new robot" (GR00T #252), custom-robot deployment questions (BeyondMimic #61), per-joint actuator tuning (#66), and URDF or actuator mismatches (holosoma's G1 ankle limits; GR00T's h2.py versus h2.urdf, #259).
   • Opportunity: validated robot assets (MJCF, URDF, USD) with identified actuator models, for multiple simulators.
6. The simulator stack keeps shifting.
   • Many flagship repos still need Isaac Gym on Python 3.8 (TWIST2, HOMIE, FALCON, PBHC).
   • The Isaac Lab 2.x to 3.0 transition breaks imports (BeyondMimic #60/#70).
   • mjlab and MuJoCo Warp, Newton 1.0 and Genesis 1.x are all ACTIVE.
   • Opportunity: environments and tasks shipped for more than one backend, with sim2sim harnesses.
7. Evaluation is weak and inconsistent.
   • HumanTracker (Aug 2026) argues kinematic error metrics "miss the physical artifacts that matter most" (foot skating, contact timing).
   • Papers compare on different splits (SONIC versus Any2Track, BeyondMimic and GMT on its own unseen set; GMR on 21 LAFAN1 clips; YAHMP on 1,024 AMASS/OMOMO motions).
   • Users run their own disturbance benchmarks (GR00T #269).
   • Opportunity: a neutral benchmark-plus-data product with held-out, licensed test motions.
8. Sim2real reliability is still the hardest part.
   • ASAP's learned delta actions exist because simulated and real dynamics diverge.
   • BeyondMimic names an unvalidated "planning-control gap."
   • Users report T1 standing sway (holosoma), SONIC sim2sim falls (#233, #270) and G1 over-temperature (#247).
   • Opportunity: system-identification data and actuator or thermal models, and domain-randomization packs tuned per robot.
9. Code and data are often only partly open.
   • Checkpoint-only releases: GMT, AMO, Humanoid-GPT (training code TODO).
   • SONIC's training code came 2 months after its weights, without the full corpus.
   • Non-commercial licenses on key projects: human2humanoid, PBHC, HOMIE, BFM-Zero, ReactiveBFM.
   • Commercially usable building blocks: SONIC (Apache plus OML), holosoma, GMR, mjlab, ProtoMotions, MimicKit, rsl_rl.
10. Compute is expensive for generalists.
    • SONIC took 9k GPU-hours; ProtoMotions used 4× A100 for 12 h (AMASS) or 24× A100 (BONES); holosoma's tracking used 4× L40s.
    • Locomotion alone is now cheap: 15 minutes on one RTX 4090 (holosoma).
    • Opportunity: pretrained, licensed base trackers per robot that customers fine-tune.

Every GitHub repo referenced (stars fetched 2026-10-02)

• NVlabs/GR00T-WholeBodyControl — ~3k (from /releases)
• HybridRobotics/whole_body_tracking — 2.4k
• HybridRobotics/motion_tracking_controller — 540
• mujocolab/mjlab — ~3k
• amazon-far/holosoma — 1.7k
• YanjieZe/GMR — 2.4k
• YanjieZe/TWIST — 818
• amazon-far/TWIST2 — 867
• NVlabs/ProtoMotions — 2.4k (main page) / 2.1k (releases page)
• xbpeng/MimicKit — 2.3k
• leggedrobotics/rsl_rl — 3k
• leggedrobotics/legged_gym — 3.1k
• unitreerobotics/unitree_rl_gym — 3.5k
• unitreerobotics/unitree_rl_lab — 1.3k
• roboterax/humanoid-gym — 2.1k
• BoosterRobotics/booster_gym — 304
• BoosterRobotics/booster_train — 56
• LeCAR-Lab/HumanoidVerse — 470
• LeCAR-Lab/ASAP — 2.1k
• NVlabs/HOVER — 758
• LeCAR-Lab/human2humanoid — 1.1k
• ZhengyiLuo/PHC — 1.3k
• zixuan417/humanoid-general-motion-tracking — 439
• LeCAR-Lab/BFM-Zero — 715
• InternRobotics/OpenHomie — 619
• OpenTeleVision/AMO — 379
• LeCAR-Lab/FALCON — 435
• TeleHuman/PBHC — 1.1k
• hongsukchoi/VideoMimic — 831
• LeCAR-Lab/HDMI — 664
• MarkFzp/humanplus — 849
• GalaxyGeneralRobotics/Humanoid-GPT — 457
• GalaxyGeneralRobotics/HumanTracker — 32 (via HF papers page, not GitHub)
• nvidia-isaac/WBC-AGILE — 346
• fan-ziqi/robot_lab — 1.9k
• robfiras/loco-mujoco — 1.5k
• carlosferrazza/humanoid-bench — 799
• google-deepmind/mujoco_playground — 2.2k
• kscalelabs/ksim — 225
• hucebot/yahmp — 12
• OpenDriveLab/EgoHumanoid — 218
• zjwzcx/ReactiveBFM — 52
• stack-of-tasks/pinocchio — 3.8k
• loco-3d/crocoddyl — 1.3k
• Simple-Robotics/aligator — 317
• stack-of-tasks/tsid — 349
• leggedrobotics/ocs2 — 1.5k
• RobotLocomotion/drake — 4.2k
• ihmcrobotics/ihmc-open-robotics-software — 325
• google-deepmind/mujoco_mpc — 1.7k
• kevinzakka/mink — 1.5k
• stephane-caron/pink — 851
• Rhoban/placo — 353
• qiayuanl/legged_control — 1.8k
• loongOpen/OpenLoong-Dyn-Control — 347
• ros-controls/ros2_control — 973
• isaac-sim/IsaacLab — 8.2k
• newton-physics/newton — 5.6k
• google-deepmind/mujoco_warp — 1.5k
• Genesis-Embodied-AI/Genesis — 30k
• ubisoft/ubisoft-laforge-animation-dataset — 1.6k
• lijiaman/omomo_release — 292
• Not fetched (mentioned only): kscalelabs/kinfer, amazon-far/PRISM, OpenDriveLab/WholebodyVLA, LeCAR-Lab/SoFTA, liuchenxu125/GMR (fork), GuilhermeAsura/humanoid_repos_eval