SuperMimic: 1 GPU, 3 hours, 129,514 motions
A new humanoid control method claims it beats a 128-GPU, 7-day setup using one GPU for three hours. Pieter Abbeel says it works because a world model is faster than a regular simulator.
Shown this week
SuperMimic trains humanoid motion on one GPU in three hours
Qiayuan Liao says SuperMimic cuts the GPU-hours needed for humanoid full-body control by more than 1000x. Across 129,514 motions, he says it beats the previous state of the art, which took 128 GPUs and 7 days, using one GPU for three hours.
Now it's superfast (and cheap, just one GPU) to train a single policy on *all* the motions.
New benchmark runs coding agents across 1,090 embodied tasks
Manling Li released the Embodied First Exam, which she says runs coding agents on 1,090 embodied tasks across 31 simulation suites and 44 robots from one line of code. Her post lists spatial-constraint puzzles, multi-robot Olympic sports, whole-body control, and real-time evaluation where the world does not pause while the agent thinks.
Teleop and other arguments
Skeptic says Unitree's car-lifting H2 was helped by a jack
Unitree posted its H2-X "Kung Fu Legend Guardian" with a claimed 8 hours of endurance, and Humanoid Scott picked apart the car-lifting clip. He says the video is real but the car was already on a jack, pointing to the wheel sitting at lowest extension and front shocks that look frozen in place.
Video is real, but car is clearly already on a jack, as wheel is at lowest extension, and jacked during “lift