Event
JARVIS releases visual manual-guided assembly benchmark
1 reports · 1 sources
The JARVIS research team released JARVIS, a method-agnostic visual manual-guided autonomous assembly benchmark, together with the accompanying embodied agent JARVIS-agentic, which can execute without additional language instructions. The benchmark contains 236,679 LEGO assembly tasks and 8,987 scenes, and additionally includes the JARVIS-v1 suite of 50 selected tasks for interactive robot evaluation. JARVIS-agentic composes robot primitives into executable programs and uses execution feedback for verification and recovery, without requiring task-specific action trajectories or model parameter updates. Across nine multimodal and reasoning models, task-oriented primitives improved the average assembly score; however, performance drops sharply as assembly duration increases, and execution errors remain difficult to correct.
Report timeline
- JARVIS: A Benchmark and Agentic Method for Autonomous Robotic Assembly from Visual Manuals
The paper introduces JARVIS, a method-agnostic benchmark for autonomous execution guided by visual manuals without additional language instructions, comprising 236,679 LEGO assembly tasks across 8,987 scenes, plus a curated JARVIS-v1 evaluation suite of 50 tasks.
arXiv cs.RO ·