SuperMimic trains humanoid tracker in 3 hours on one GPU, beating 128-GPU 7-day SoTA
daibond_alpha · x · 2026-11-10
A technical thread from daibondalpha breaks down SuperMimic, which trains a humanoid tracker surpassing a SoTA baseline (7 days on 128 GPUs) in just 3 hours on a single GPU.
The key is differentiable planning: it differentiates predicted return through a learned dynamics model via BPTT and bootstraps with a terminal critic. The accompanying figure contrasts PPO's score-function update (weighing sampled actions by advantages) with the pathwise BPTT update, which chains local sensitivities (policy, action, dynamics, value gradients) through the imagined trajectory while keeping dynamics weights fixed.
The estimator itself isn't new — it matches the pathwise value gradient of SVG (Heess et al., 2015) and Dreamer v1 (Hafner et al., 2020); the win comes from stable and effective gradients.
Related event: SuperMimic Trains SOTA Humanoid Tracker in 3 Hours on a Single GPU(5 posts)→
More from Embodied
- LingBot-Touch lifts VLA success rate from 45.6% to 72.5%, pretrained on 4,983 hours of tactile data — CyberRobooo · 2026-11-10
- How LingBot-Touch works: one shared Haptic Transformer unifies magnetic, piezoresistive and visuotactile signals — CyberRobooo · 2026-11-10
- Robbyant releases LingBot-Touch, a tactile model giving robot hands human-like touch sensing — CyberRobooo · 2026-11-10
- Claude Fable 5.2 scores 64.9% on RoboDojo robot control, dominating rivals — burny_tech · 2026-11-10
- Dyna Robotics shows RL vs IK head-to-head on wheeled robot, unveils workflow agent Dyna-2.1 — chris_j_paxton · 2026-11-10
- AI researcher Chris Paxton comes away positive from a Tesla Cybercab ride — chris_j_paxton · 2026-11-10