Long-Horizon Multi-Object Rearrangement via Distilled Tree Search over Continuous Actions

Key points

• Despite recent advances in policy learning, long-horizon manipulation remains difficult because learned policies must avoid compounding errors while preserving future feasibility.1
• This raises a central question: how can long-horizon reasoning from search be transferred into policies in continuous action spaces?1
• DISCO uses progressive-widening MCTS in physics simulation to generate long-horizon plans.1
• Across simulated and real-world rearrangement tasks, DISCO outperforms behavior cloning, reinforcement learning, and search-based baselines.1

Sources (1)

• [1]Long-Horizon Multi-Object Rearrangement via Distilled Tree Search over Continuous Actions
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 19, 04:50 AM

  To this end, we propose DISCO: Distilled Tree Search over Continuous Actions.

  Despite recent advances in policy learning, long-horizon manipulation remains difficult because learned policies must avoid compounding errors while preserving future feasibility.

Extractive summary: sentences quoted from the sources.