UC Berkeley's SuperMimic Reduces Whole-Body Control Training from Multi-GPU Clusters to a Single GPU
机器人的训练成本被 Supermimic 打下来了?
UC Berkeley researchers present SuperMimic, a whole-body control framework based on a single differentiable world model that reduces training costs for large-scale human motion imitation from multi-GPU clusters running days to a single NVIDIA H100 GPU running about 3 hours. The method leverages shared rigid body dynamics across diverse human motions to enable analytical gradients and fast imagination, achieving zero failure rate on the training set after 9.8 hours on two GPUs in tests involving 129,514 human motions (approx. 264 hours). Accepted as an oral presentation at CoRL 2026, SuperMimic challenges the prevailing reliance on model-free reinforcement learning and massive compute clusters for WBC scaling.
Source: Chinese robotics — Research news discovery · Read original article ↗
English translation unavailable. Choose the original article to read it.
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
GPU hours
≈21,000
Summed sensor-hours
View original evidence
3、英伟达团队的「SONIC」作为此前最优的动作追踪器,为了让机器人掌握 12.9 万种不同的人类动作,需要调用 128 张 NVIDIA L40 GPU 在集群上进行长达 7 天的训练,累计消耗约 21,000 个 GPU 小时。
Open source S19Reported duration
≈3 hours
View original evidence
[1-2] 4、Supermimic 引起的热议在于,该工作将处理大规模人类动作跟踪策略的训练成本从 128 张 GPU 运行数天,压缩至单张 GPU 运行约 3 小时。
Open source S20GPU cards
128
View original evidence
3、英伟达团队的「SONIC」作为此前最优的动作追踪器,为了让机器人掌握 12.9 万种不同的人类动作,需要调用 128 张 NVIDIA L40 GPU 在集群上进行长达 7 天的训练,累计消耗约 21,000 个 GPU 小时。
Open source S19GPU cards
1–2
View original evidence
[1-2] 4、Supermimic 引起的热议在于,该工作将处理大规模人类动作跟踪策略的训练成本从 128 张 GPU 运行数天,压缩至单张 GPU 运行约 3 小时。
Open source S20hours
7
Unique elapsed hours
View original evidence
3、英伟达团队的「SONIC」作为此前最优的动作追踪器,为了让机器人掌握 12.9 万种不同的人类动作,需要调用 128 张 NVIDIA L40 GPU 在集群上进行长达 7 天的训练,累计消耗约 21,000 个 GPU 小时。
Open source S19Reported duration
9.8 hours
human action segments
129,514
View original evidence
在包含 129,514 段人类动作(约 264 小时)的测试中,SuperMimic 在单张 NVIDIA H100 GPU 上运行 3 小时的归一化回报率超过了 SONIC 的公开检查点,且在失败率上表现更优。
Open source S22
- paper: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
3、英伟达团队的「SONIC」作为此前最优的动作追踪器,为了让机器人掌握 12.9 万种不同的人类动作,需要调用 128 张 NVIDIA L40 GPU 在集群上进行长达 7 天的训练,累计消耗约 21,000 个 GPU 小时。
Open source S19
[1-2] 4、Supermimic 引起的热议在于,该工作将处理大规模人类动作跟踪策略的训练成本从 128 张 GPU 运行数天,压缩至单张 GPU 运行约 3 小时。
Open source S20
在双卡运行 9.8 小时后,其在训练集内的测试动作达到零失败。
Open source S23
在包含 129,514 段人类动作(约 264 小时)的测试中,SuperMimic 在单张 NVIDIA H100 GPU 上运行 3 小时的归一化回报率超过了 SONIC 的公开检查点,且在失败率上表现更优。
Open source S22
Source:Chinese robotics — Research news discovery · finance.sina.com.cn