What is motion tracking and why did it spread so fast?
In motion tracking the policy has a simple job: at every control step, bring the robot as close as possible to the next pose of the reference motion. The reward is computed from the difference between the robot's joint angles and body positions and the reference. The hard part is reproducing human motion in a way that fits the robot's mass, joint limits and motor power without falling.
The pipeline has four steps:
- Human motion is recorded with motion capture or estimated from video.
- The motion is retargeted to the robot's skeleton. NVIDIA's tool for this step on the G1 is covered in our SOMA Retargeter guide.
- A policy learns to follow the reference in simulation with thousands of parallel robots.
- The policy is tested first in another simulator (sim2sim) and then on the real robot.
In the SONIC paper NVIDIA describes motion tracking as a task that can be trained at scale. Instead of designing a separate reward for each behavior, one tracking objective is used and the variety of skills comes from the data. The methods of the past year are built on this idea.
How the field developed: a timeline
The dates below come from the first arXiv versions of the papers and the creation dates of the GitHub repositories.
- 2020: Ubisoft La Forge released the LAFAN1 motion dataset with a paper presented at SIGGRAPH 2020. Most of today's G1 comparisons use 40 clips from this data.
- September 2024: NVIDIA opened the ProtoMotions repository. The framework is now in its third version and includes a G1 deployment pipeline.
- June 2025: The mjlab repository was opened. Its paper (arXiv 2601.22074) followed on 29 January 2026.
- 11 August 2025: Researchers from UC Berkeley and Stanford released the BeyondMimic paper (arXiv 2508.08241) and its motion tracking code.
- 11 November 2025: NVIDIA released the GEAR SONIC paper (arXiv 2511.07820).
- 16 March 2026: The BONES-SEED dataset was opened. According to the GR00T-WholeBodyControl README it holds more than 142,000 motions and about 288 hours of recordings.
- 20 October 2026: The UC Berkeley team released the SuperMimic paper (arXiv 2610.28480), accepted as an oral presentation at CoRL 2026.
- 21 October to 15 November 2026: Open-source repositories built on the SuperMimic idea followed one another: supermimic_vibed (21 October), InstantTrack (29 October), BlinkMimic (2 November), XMimic (7 November, paper on 16 November) and MacMimic (15 November). SuperMimic's official code was opened in the HybridRobotics repository on 10 November.
Comparing the methods
The first table shows who proposed what, the second shows the infrastructure and the results. The time column is each project's own measurement. The projects were measured with different data and targets, so rows cannot be compared directly for speed.
| Project | Who | Year | Core of the method |
|---|---|---|---|
| BeyondMimic | UC Berkeley (Hybrid Robotics) and Stanford | 2025 | Motion tracking with one MDP and shared hyperparameters, with a task-guided latent diffusion model on top |
| GEAR SONIC | NVIDIA (GEAR) | 2025 | Scaling motion tracking: 700 hours of motion capture data, networks from 1.2M to 42M parameters |
| ProtoMotions | NVIDIA (NVlabs) | 2024 | GPU-accelerated training framework for simulated humans and humanoids, general G1 tracking policy |
| mjlab | Zakka, Liao, Yi, Le Lay, Sreenath, Abbeel | 2025 | Training framework that combines an Isaac Lab style manager API with MuJoCo Warp |
| SuperMimic | UC Berkeley | 2026 | One differentiable world model for all motions, policy trained by backpropagation through short imagined rollouts |
| supermimic_vibed | Chong Zhang (ETH Zurich) | 2026 | Independent, unofficial reimplementation of SuperMimic |
| InstantTrack | Huy-Quang-Dao | 2026 | GPU loop engineering on top of SuperMimic, EMA policy, Muon optimizer and failure-focused resets |
| BlinkMimic | Fuyi Chen | 2026 | Semi-implicit world model, mirror augmentation and evaluation-guided resets on top of InstantTrack and supermimic_vibed |
| XMimic | Nico Bohlinger, Jan Peters | 2026 | BlinkMimic's algorithm rebuilt in JAX, with each training iteration compiled into one XLA program |
| MacMimic | air23zj | 2026 | BlinkMimic's algorithm with XMimic's simulator and learner overlap, MLX on Apple silicon |
| Project | Infrastructure | Hardware and time (per source) | Validation on robot | Code licence |
|---|---|---|---|---|
| BeyondMimic | Isaac Lab 2.1.0, RSL-RL PPO; code now lives in mjlab | Tracking inference on the robot CPU under 1 ms | Real G1 (paper supplement) | MIT |
| GEAR SONIC | Isaac Lab training, C++ and TensorRT deployment | 21k GPU hours | Real G1, 124 sequences (NVIDIA) | Apache-2.0, weights under NVIDIA Open Model License |
| ProtoMotions | IsaacGym, IsaacLab, Newton, MuJoCo | 12 hours on 4 A100s for AMASS, 24 A100s for BONES-SEED | Real G1, zero-shot transfer (NVIDIA) | Apache-2.0, models under NVIDIA Open Model License |
| mjlab | MuJoCo Warp | NVIDIA GPU for training, evaluation on macOS | Training framework with a ready G1 tracking task | Apache-2.0 |
| SuperMimic | MuJoCo Warp, code built on mjlab | About 3 hours on one H100, 9.8 hours on two GPUs for zero failures | Real G1 (paper) | Apache-2.0 |
| supermimic_vibed | Isaac Lab 3, Newton with the MuJoCo-Warp solver | About 0.95 hours on an RTX PRO 6000 Blackwell | Simulation (40 LAFAN1 clips) | MIT |
| InstantTrack | mjlab 1.6.0, MuJoCo Warp | About 3 minutes on an RTX 5090 | Simulation (40 LAFAN1 clips) | MIT |
| BlinkMimic | mjlab, MuJoCo Warp | 2.9 minutes to target on an RTX 4090 | Simulation (40 LAFAN1 clips) | MIT |
| XMimic | JAX, MuJoCo Warp through MJX | 0.86 minutes to target on 4 H100s | Simulation (40 LAFAN1 clips) | MIT |
| MacMimic | MuJoCo on the CPU, MLX on the GPU | Mean of 21.6 minutes to target on a MacBook Pro M4 Pro | Simulation (40 LAFAN1 clips) | MIT |
SuperMimic and its derivatives: speed through a world model
To understand the speed-up of the last two months, start with SuperMimic. In the classic approach the policy improves by trial and error in the simulator with a reinforcement learning method such as PPO. SuperMimic instead builds one differentiable world model that learns the robot's dynamics from transitions logged in the simulator. The policy gets analytic gradients by backpropagation through time (BPTT) along short rollouts imagined inside that model.
According to the paper, SuperMimic was trained on 129,514 motions filtered from BONES-SEED. On a single H100 it surpassed SONIC's normalized return in about 3 hours. The authors report that the SONIC baseline was trained for 7 days on 128 GPUs and that they reduced GPU hours by more than 1000 times. After 9.8 hours of training on two GPUs the number of failed motions dropped to zero.
The open-source repositories that followed focused on speeding up the same idea on a 40-clip LAFAN1 benchmark. The shared target is a normalized return of 0.82 with no failed clips. The chain went like this:
- supermimic_vibed: Chong Zhang's reimplementation reached 0/40 failed clips and a return of 0.842. The README states openly that the code was written in one hour with an AI coding assistant and has not been reviewed.
- InstantTrack: started from the mjlab port of supermimic_vibed. The author's own SuperMimic reimplementation took 1.50 hours on the same machine, while InstantTrack reached the target in about 3 minutes.
- BlinkMimic: combined the two and held the target in 2.9 minutes on one RTX 4090, averaged over 3 seeds. A standard run is about 8 minutes for 3,500 iterations.
- XMimic: changed the machinery rather than the method. The loop compiled with JAX came down to 0.86 minutes on 4 GPUs and 2.47 minutes on one GPU. The paper also presents XSoma for rough-terrain locomotion on the same infrastructure.
- MacMimic: moved the pipeline to an Apple silicon laptop. MuJoCo simulates 4,096 robots on the CPU while the world model, policy and critic train on the GPU with MLX.
The READMEs of InstantTrack, XMimic and supermimic_vibed state that they are not affiliated with the SuperMimic authors. The results of these repositories are in simulation, and the real-robot result is in SuperMimic's own paper.
From laptop to lab: which hardware to start with
Motion tracking is no longer tied to one class of hardware. The table below puts the hardware and time reported by each source side by side. You can start by understanding the method on a laptop and plan a path that ends with a real G1 in the lab.
| Level | Hardware (per source) | Project | Result reported by the source |
|---|---|---|---|
| Simulation on a laptop | MacBook Pro, M4 Pro, 48 GB | MacMimic | Mean of 21.6 minutes to target, full 2,000-iteration run about 30 minutes |
| One consumer GPU | NVIDIA RTX 4090 or RTX 5090 | BlinkMimic, InstantTrack (mjlab) | About 3 minutes to target on 40 LAFAN1 clips |
| Workstation GPU | RTX PRO 6000 Blackwell or one H100 | supermimic_vibed, SuperMimic | About 1 hour on LAFAN1, about 3 hours on BONES-SEED |
| Multi-GPU server | 4 H100s or 4 to 24 A100s | XMimic, ProtoMotions | 0.86 minutes on LAFAN1, 12 hours on AMASS |
| Large scale and a real G1 | GPU cluster and a Unitree G1 | GEAR SONIC | 21k GPU hours, real-G1 evaluation with 124 sequences |
MacMimic's speed comparison is on the same Mac. The repository's MLX port of BlinkMimic reached the target in 44 minutes, so MacMimic is 2.04 times faster. The README notes that the comparison uses its own MLX ports rather than the original CUDA code, and that BlinkMimic reaches slightly higher precision at the target (0.847 against 0.822). mjlab supports macOS for evaluation and needs an NVIDIA GPU for training.
The real-robot stage needs a Unitree G1 EDU. For teams that plan the lab setup as one package, we prepared the G1 research platform page.
Data and licence notes
Most methods ship with openly licensed code, but the training data has its own licence, often a narrower one. If you plan a commercial product, read the data row before the code row.
| Data | Content | Licence | Projects using it |
|---|---|---|---|
| LAFAN1 (Ubisoft La Forge) | 5 actors, 77 sequences, about 4.6 hours at 30 fps | CC BY-NC-ND 4.0 (non-commercial, no derivatives) | BeyondMimic, InstantTrack, BlinkMimic, XMimic, MacMimic, supermimic_vibed |
| G1 retargeting of LAFAN1 (Hugging Face, lvhaidong/LAFAN1_Retargeting_Dataset) | Motions for the G1, H1 and H1_2 | The dataset card states the LAFAN1 licence, code under MIT | Same projects |
| AMASS (Max Planck Institute) | Unified motion capture archive | Non-commercial scientific research licence, no redistribution | ProtoMotions |
| BONES-SEED (Bones Studio) | More than 142,000 motions, about 288 hours, SOMA and G1 formats | BONES-SEED License: academia and startups with annual revenue under USD 1 million, a separate commercial licence for other organisations | GEAR SONIC, SuperMimic, ProtoMotions |
All repositories that use LAFAN1 stress the same point: the data is not stored in the repository and is downloaded at setup. Checkpoints and videos derived from this data are outside the MIT licence and are for non-commercial research. The G1 robot description is licensed under BSD-3-Clause. The BONES-SEED terms are explained one by one in our BONES-SEED guide.
From simulation to a real G1
A policy that succeeds in simulation is moved to a real G1 in two stages. First comes a sim2sim test in a physics engine other than the one used for training, usually MuJoCo. Then the same inference code is connected to the robot. The three pipelines that publish real-G1 results follow different routes:
- BeyondMimic: HybridRobotics' motion_tracking_controller repository is a C++ inference pipeline on ROS 2 Jazzy and legged_control2 that runs the ONNX policy on the CPU. The same package holds a MuJoCo sim2sim launcher and a real-robot launcher over Ethernet. The policy is started with the Unitree joystick, and the B button is the emergency stop.
- GEAR SONIC: deployment uses C++ and TensorRT, and the Jetson Orin on the G1 needs JetPack 6 and TensorRT 10.7. Details are in our GEAR SONIC guide.
- ProtoMotions: the tracker trained on BONES-SEED is exported to ONNX, tested in MuJoCo and connected to a real G1 through the RoboJuDo framework. The steps are in our ProtoMotions guide.
For repositories that report results only in simulation, this stage is up to the user. The MacMimic README states plainly that the policy has not been run on a physical robot. Before moving a policy from these repositories to the robot, run a sim2sim test in MuJoCo and protect the robot with a gantry or support. For Unitree's own reinforcement learning workflow see our RL Gym page, and for Turkish documentation of the MuJoCo simulator see our unitree-mujoco-turkce repository.
Frequently asked questions
Is motion tracking the same as imitation learning?
They are close relatives. In motion tracking the policy follows a reference motion at joint level and is trained with reinforcement learning in simulation. Imitation learning usually learns a task policy from demonstrations collected through teleoperation. For that second workflow see our LeRobot guide.
Which project can I start with before buying a robot?
If you have an Apple silicon laptop, MacMimic is a good start. With a single RTX 4090 or 5090, BlinkMimic or InstantTrack fit well. All three learn the 40 LAFAN1 clips in simulation in anything from a few minutes to half an hour.
Which methods have published results on a real G1?
BeyondMimic, SuperMimic, GEAR SONIC and ProtoMotions. The results are in the respective papers and on NVIDIA's project pages.
Can I use a policy trained on LAFAN1 in my product?
LAFAN1 is licensed under CC BY-NC-ND 4.0. The repositories restrict checkpoints derived from this data to non-commercial research. For commercial use, retrain on data with a suitable licence and talk to your legal team.
How does SONIC differ from SuperMimic?
SONIC scales motion tracking with large data and thousands of GPU hours and ships with a real-G1 deployment stack. SuperMimic brings training down to a single GPU on the same data family with a differentiable world model. According to its paper, the policy was transferred to a real G1.
Is there a Turkish resource?
Yes. The unitree-g1-hareket-taklidi-arastirma repository collects this article's comparison tables, data licences and source list in Turkish.
Motion imitation work with the Unitree G1 in Turkey
As the Unitree Robotics Official Distributor for Turkey, Robotlar.org works with university and company teams doing research on the G1 family. We collected this article's tables and source list in Turkish in the unitree-g1-hareket-taklidi-arastirma repository. For the full picture of NVIDIA's projects see our NVIDIA humanoid motion family hub guide, and for the GR00T workflow our Unitree G1 and Isaac GR00T guide.
We decide together with you in a discovery call at which level your motion imitation work should start. Some teams begin by understanding the method on a laptop, others move straight to sim2real work on a real robot with a G1 EDU. For configuration and pricing use the G1 EDU quote page or the corporate quote form. Our G1 research platform page covers lab setups, and Unitree education solutions covers courses and training programmes.
Sources
- arXiv 2508.08241, BeyondMimic: From Motion Tracking to Versatile Humanoid Control via Guided Diffusion
- BeyondMimic project page
- HybridRobotics/whole_body_tracking (BeyondMimic motion tracking code)
- HybridRobotics/motion_tracking_controller (sim2sim, sim2real)
- arXiv 2610.28480, SuperMimic: Scaling Humanoid Whole-Body Control with One Differentiable World Model
- SuperMimic project page
- HybridRobotics/supermimic
- zita-ch/supermimic_vibed
- Huy-Quang-Dao/InstantTrack
- fuyichen0725/BlinkMimic
- arXiv 2611.20452, XMimic and XSoma: Learning Humanoid Motion Tracking and Locomotion in Under a Minute
- nico-bohlinger/xmimic
- MacMimic project page
- air23zj/macmimic
- arXiv 2601.22074, mjlab: A Lightweight Framework for GPU-Accelerated Robot Learning
- mujocolab/mjlab
- arXiv 2511.07820, SONIC: Supersizing Motion Tracking for Natural Humanoid Whole-Body Control
- NVlabs/GR00T-WholeBodyControl (GEAR SONIC)
- nvidia/GEAR-SONIC, Hugging Face
- NVlabs/ProtoMotions
- Ubisoft La Forge, LAFAN1
- LAFAN1_Retargeting_Dataset, Hugging Face
- AMASS license
- BONES-SEED, Hugging Face
- unitree-g1-hareket-taklidi-arastirma

