SJTU's VaM lets robots learn online from mistakes, 95%+ success over 4-hour autonomous runs
chris_j_paxton · x · 2026-10-26
Shanghai Jiao Tong University introduces Value as Memory (VaM), an online value-guided post-training framework for general robot policies. Recent value predictions are cached as task-progress memory alongside visual and proprioceptive observations, letting policies distinguish visually similar states (lost contact vs. successful re-grasp) and recover from mistakes. Training spans four offline-to-online stages, reusing demonstrations, off-policy human corrections, and on-policy rollouts with concurrent data collection and updates. VaM achieves 95%+ task success over 4 hours of continuous autonomous box assembly, also validated on laundry folding and magnetic bus assembly. Paper, code, and project page are public.
More from Embodied
- WanGP v18.81 adds Tencent's Hy-Nomas world model, playable on 8GB VRAM — cocktailpeanut · 2026-10-27
- MS-Human-700: open-source human model with 700 muscle actuators lands in MuJoCo — yisongyue · 2026-10-27
- NVIDIA open-sources FLORA, a robotics data engine ~125x cheaper than GPT-6 Astra — nvidia · 2026-10-27
- Embodied AI startups need a different operational discipline, investor says — ___Mufasaa · 2026-10-27
- π0.5 study: raising demos from 10 to 438 lifts Place Ball in Bowl success from 7% to 83% — stingning · 2026-10-27
- GPT-6 Astra controls dexterous hands zero-shot at just 5.2% success in 500 trials — stingning · 2026-10-27