CUHK's Orchestra robot framework completes 7.35-step subtask sequences
CUOHK · hf · 2026-10-14
Researchers from CUHK, Xiaomi and NUDT introduce Orchestra, an agent harness for long-horizon robotic manipulation: it decomposes multimodal instructions into ordered subtasks, tracks progress, and updates object locations via a shared context, with a separately trained progress evaluator deciding when to advance. They also release the RMBench-Compose benchmark. Orchestra hits 68.22% mean success across nine RMBench tasks (+26.22 points over Mem-0), achieves an Average Success Length of 7.35 subtasks (2.46x HarnessVLA) on RMBench-Compose, and improves real-world ASL by 25% over the base π0.5 policy. Code and project page are open-sourced.
More from Embodied
- Ukrainian FPV Drone Sim Leaves Early Access, Players Generate Training Data — generativist · 2026-10-14
- Menlo Research teases mystery robotics kit build in new photo — victormustar · 2026-10-14
- China's Hesai launches spatial camera to boost AI-powered robotics — pstAsiatech · 2026-10-14
- Nvidia's TacitVLA matches CoT performance at up to 3.4x lower latency — arankomatsuzaki · 2026-10-14
- $3,500 Flourish 1 Home Robot Uses a Raspberry Pi and Learns Chores From Your Phone — Matt Wolfe · 2026-10-14
- SlashVibe, a physical controller for AI agents, hits Kickstarter October 27 — sujingshen · 2026-10-14