SuperMimic trains humanoid tracker in 3 hours on one GPU, beating 128-GPU SoTA baseline
daibond_alpha · x · 2026-11-10
STACX Lab's blog demystifies why SuperMimic trains a humanoid tracker in 3 hours on a single GPU, surpassing a SoTA baseline trained 7 days on 128 GPUs: differentiable planning with stable, effective gradients.
Key points:
- SuperMimic differentiates predicted return through a learned dynamics model (BPTT) and bootstraps with a terminal critic — the estimator comes from SVG's pathwise value gradient (Heess et al., 2015) and DreamerV1.
- The field abandoned this route (DreamerV2/V3 went REINFORCE-style, robotics converged on PPO), but the control interface changed the math: 50 Hz PD position control pulls errors back each step, keeping BPTT gradients bounded, while direct torque control makes them explode.
- Once gradients are stable, efficiency follows: pathwise gradient SNR is 30× higher than score-function, 64× fewer imagined steps than PPO on the same rollouts, and a terminal critic keeps rollouts short.
Related event: SuperMimic Trains SoTA Humanoid Tracker in 3 Hours on a Single GPU(3 posts)→
More from Embodied
- Calibrex, an open-source robot sensor calibration checker, hits 50 GitHub stars — rsasaki0109 · 2026-11-10
- SpaceX says Grok detects aerospace foreign object debris with 99.9% accuracy vs humans' 94% — elonmusk · 2026-11-10
- HermesEnabler2 confirmed working on Apple Watch Series 10, unlocking Hermès watch faces — gregbarbosa · 2026-11-10
- Fable 5.2 hits 64.9% on RoboDojo, beating purpose-built robot AI systems — CyberRobooo · 2026-11-10
- Caltech robot builds a structure in orbit autonomously, a first for space construction — GregCook2011 · 2026-11-10
- Robot paints original watercolors and ships them with telemetry — a one-man robot art club — repligate · 2026-11-10