DISCO: Distilled Tree Search Beats RL and BC on Long-Horizon Robot Rearrangement, CoRL 2026 Oral
stanfordnlp · x · 2026-10-26
A team including Kallol Saha, Yejin Choi, Maxim Likhachev and David Held proposes DISCO (Distilled Tree Search over Continuous Actions). It runs progressive-widening MCTS inside physics simulation to plan long-horizon multi-object rearrangement: a learned policy proposes continuous actions while a learned value function scores expected task progress. With GPU-parallelized simulation, DISCO periodically distills MCTS-generated plans back into the policy-value model, so the final model deploys directly without simulator-based search at test time. Across simulated and real-world rearrangement tasks, DISCO outperforms behavior cloning, reinforcement learning, and search-based baselines. Accepted at CoRL 2026 as an Oral.
More from Embodied
- Why ultrasound is the goldilocks zone for brain-computer interfaces — burny_tech · 2026-10-26
- A tiny dashcam is actually driving the car — hands off, eyes on — k7agar · 2026-10-26
- Hobbyist completes DIY bipedal robot lower body with KONDO servos and WIT IMU — ___Mufasaa · 2026-10-26
- Amazon Robotics CTO: at 1M robots deployed, 99% reliability still means millions of errors — ___Mufasaa · 2026-10-26
- Early hands-on: "It's really good" says user after trying the Duo device — jdluk87 · 2026-10-26
- Tesla Cybercab convoy filmed driving in formation on highway — yunta_tsai · 2026-10-26