SuperMimic presents a method for training a single humanoid policy to track a large library of whole-body motions, arguing that compute is no longer the main bottleneck when learning uses a differentiable world model. Instead of differentiating through a contact simulator, it gathers transitions from MuJoCo Warp, fits a learned model of the robot’s dynamics, and updates the policy through imagined rollouts in that model. The world model and policy train together; the full loop, including simulator data collection, model fitting, and policy updates, takes three hours on one NVIDIA H100 to exceed the released SONIC checkpoint’s normalized return of 0.90. The comparison is striking: SONIC used 128 L40 GPUs for seven days, or about 21,000 GPU-hours, while SuperMimic used three GPU-hours and 2.3 times less motion data.
On a benchmark of 129,514 motions, SONIC’s checkpoint fails 179, including all 44 flips and all 22 side rolls. SuperMimic fails 25 after three hours on one GPU and none after 9.8 hours on two GPUs. These counts measure coverage of training motions, not generalization. The approach also runs on a physical Unitree G1: one policy handles the demonstrated motions without per-motion tuning. The authors attribute its efficiency to a single model amortizing the motion library, imagined rollouts running over ten times faster than MuJoCo Warp at practical batch sizes, and first-order gradients needing far fewer imagined steps than PPO. The comparison is not identical in hardware or data, but the reported differences do not account for the large GPU-hour gap.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.