ParallelWorld uses multi-step simulation for embodied planning
Researchers from Tsinghua University and the National University of Singapore proposed ParallelWorld, which explores candidate action paths, compares predicted observations, and executes only the first action before replanning. According to 机器之心, the paper reports overall task success rising from 56% to 83%; ParallelWorld beat sequential Active Exploration on 26 of 28 ESI-Bench tasks and exceeded the direct baseline in all eight LIBERO-Pro settings. The results also include regressions on rigid-containment and geometric-configuration tasks, while adaptive search took about 5.8 times as long as direct execution.
ParallelWorld provides a concrete approach to spending test-time computation on high-level embodied planning, with reported gains on spatial understanding and robot manipulation benchmarks.
机器之心 2026-10-26 14:13 北京
ParallelWorld 将多步预演引入具身智能体的高层规划,将任务成功率从 56% 提升至 83%。
机器人动手前,能不能先看一看不同选择会带来什么后果?ParallelWorld 将多步预演引入具身智能体的高层规划,将任务成功率从 56% 提升至 83%。
把桌上的两把摩卡壶放到炉灶上,看起来只需要 “拿起来,放过去”。但对机器人来说,先搬哪一把、从哪里抓、放下后是否影响下一步,都可能改变任务的结局。
#ParallelWorld#Tsinghua University#National University of Singapore#HarnessVLA#Codex#ESI-Bench#LIBERO-Pro