Game changer for humanoid full-body control training
Pieter Abbeel and his team introduce SuperMimic, a method that enables fast and cost-effective training of full-body humanoid control policies. By using a faster world model and more efficient backpropagation through time, they achieve results previously requiring 128 GPUs over 7 days with just one GPU in 3 hours.
The research introduces SuperMimic, a method that significantly reduces the computational cost of training full-body humanoid control policies by leveraging a faster world model and more efficient backpropagation through time, achieving results previously requiring 128 GPUs over 7 days with just one GPU in 3 hours.
Source: Pieter Abbeel — X Robotics · Read original article ↗
Article text · Original source · English
Game changer for humanoid full-body control training. Now it's superfast (and cheap, just one GPU) to train a single policy on *all* the motions. How come:
(1) World model much faster than regular sim
(2) Backprop through time collects way more signal than policy gradients
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
GPU-hours saved
>1,000
View original evidence
【引用 @qiayuanliao】We found a way to scale humanoid control that reduces GPU-hours by more than 1000×.
Open source S5motions processed
129,514
View original evidence
Across 129,514 motions, SuperMimic surpasses the state of the art (128 GPUs, 7 days) with 1 GPU in 3 hours.
Open source S6training time
3 hours
View original evidence
Across 129,514 motions, SuperMimic surpasses the state of the art (128 GPUs, 7 days) with 1 GPU in 3 hours.
Open source S6GPU used
1
View original evidence
Across 129,514 motions, SuperMimic surpasses the state of the art (128 GPUs, 7 days) with 1 GPU in 3 hours.
Open source S6state of the art GPU-hours
128
View original evidence
Across 129,514 motions, SuperMimic surpasses the state of the art (128 GPUs, 7 days) with 1 GPU in 3 hours.
Open source S6state of the art days
7
View original evidence
Across 129,514 motions, SuperMimic surpasses the state of the art (128 GPUs, 7 days) with 1 GPU in 3 hours.
Open source S6state of the art GPUs
128
View original evidence
Across 129,514 motions, SuperMimic surpasses the state of the art (128 GPUs, 7 days) with 1 GPU in 3 hours.
Open source S6
- paper: not_reported
- code: not_reported
- weights: not_reported
- dataset: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
【引用 @qiayuanliao】We found a way to scale humanoid control that reduces GPU-hours by more than 1000×.
Open source S5
Across 129,514 motions, SuperMimic surpasses the state of the art (128 GPUs, 7 days) with 1 GPU in 3 hours.
Open source S6
We found a way to scale humanoid control that reduces GPU-hours by more than 1000×. Across 129,514 motions, SuperMimic surpasses the state of the art (128 GPUs, 7 days) with 1 GPU in 3 hours. Project: https://supermimic.github.io Demo: https://supermimic.github.io/demo/View the quoted post on X
Source：Pieter Abbeel · x.com