Open-Source SuperMimic Projects Cut Unitree G1 Motion-Tracker Training to Minutes on One GPU
- Published
- Source
- HumanoidsDaily
A UC Berkeley paper, SuperMimic, and a series of open-source projects built on it have sharply reduced the time needed to train a working motion tracker for the Unitree G1. The article reports that training has fallen from about one week on 128 GPUs to as little as three minutes on one GPU, according to Ros Shinkle, founder and CEO of Tuesday Labs and a former Boston Dynamics staff roboticist. SuperMimic uses a learned, differentiable model of robot dynamics rather than relying solely on model-free reinforcement learning.
The paper reports that SuperMimic trained for three hours on one GPU exceeded the return of a state-of-the-art tracker trained for seven days on 128 GPUs, with a GPU-hour reduction of more than 1,000 times. On a benchmark of 129,514 motions, it failed on 25, compared with 179 for the baseline; a 9.8-hour run on two GPUs had no failures. The authors also report transfer to a physical Unitree G1 across diverse and acrobatic motions. These results point to lower compute requirements for developing humanoid whole-body controllers.
The official code from UC Berkeley’s Hybrid Robotics group appeared on GitHub around November 10 under an Apache-2.0 license, after community implementations had begun. Chong Zhang’s October 21 project became a starting point for later work, while the community list awesome-supermimic tracks more than 40 implementations, extensions, and demos. The article’s account ends as InstantTrack’s subsequent additions are introduced, so their details and any later milestones are not disclosed.
Original report
After SuperMimic, Open-Source Humanoid Motion Trackers Now Train in Minutes on One GPU
HumanoidsDaily · [email protected] (Humanoids Daily Staff)This briefing is an independently written summary based on publicly available reporting and is provided for industry information and news discovery. The original report and source publication are credited and linked where applicable. RobotToday does not claim ownership of third-party source material.
Rights concerns? Contact [email protected] with the relevant URL and details. We will review the matter and take appropriate action where warranted.
More related news
- 1
- 2
- 3
- 4
- 5
- 6
- 7
- 8
- 9
- 10
- 11
-
12
ROBOTTODAY Weekly October 19 – 23, 2026 RobotToday · 3 weeks ago
Related Suppliers
Hybrid Robotics Lab, UC Berkeley
UC Berkeley lab led by Prof. Koushil Sreenath researching nonlinear control of legged and aerial robots; creator of the open-source Berkeley Humanoid Lite.
UC Berkeley Robotics and Intelligent Machines / BAIR Lab
UC Berkeley's AI and robotics research lab with 50+ faculty; known for robot learning, manipulation, and reinforcement learning research by Abbeel.
Tuesday Lab
New York-based robotics startup building delightful home tidying robots with supervised autonomy, founded by former Boston Dynamics engineers, YC W24.
Acheron Labs
Vienna-based robotics software startup building an LLM-driven control layer that lets Unitree's G1 humanoid and Go2 quadruped follow spoken commands.
Humanola
US start-up founded in 2025 offering VR teleoperation software for humanoids and robot arms and turning each operator session into robot-learning training data.
Shenzhen Yuhesen Technology Co., Ltd 深圳煜禾森科技有限公司
Shenzhen-based maker of modular AGV/AMR chassis platforms and SLAM navigation modules for logistics, agriculture and inspection robots.
Raybot Smartworks 苏州睿柏智悦机器人科技有限公司
Kunshan-based robotics firm behind the RB-A850-10 harmonic robotic-arm platform for humanoid and wheeled robot integrators.
Seoul Dynamics Inc.
Korean maker of heavy-duty physical-AI mobile robots: the RAPTOR LIFT 1.5-2 t riderless autonomous forklift and the 10-15 t TYRANNO platform for yards/sites.
Google DeepMind
Alphabet's London-based AI lab, developer of the Gemini Robotics vision-language-action models and the open-source MuJoCo physics simulator.
Noble Machines
US developer of general-purpose humanoid robots that learn industrial skills in hours via language, demonstrations, and gestures.
Shenzhen Liangyuan Innovation Technology Co., Ltd. 深圳亮源新创科技有限公司
Shenzhen embodied-AI lab led by ex-OpenAI researcher Roger Jiang, building Physical AI foundation models Light-O1, Light REACT and LightNav-0
Boardwalk Robotics
Pensacola-based humanoid robotics company developing Alex, a 19-DOF legless upper-torso humanoid for manufacturing and logistics.
Robotics needs a service framework.
RSF defines a common language for robot service capability, lifecycle operations, certification pathways, and service-provider networks.
Join the RobotToday community on LinkedIn
Daily robotics news, in-depth analysis, conference highlights, and discussions with professionals worldwide.