Event

TrySee proposes a closed-loop robot task planning pipeline without fine-tuning

1 reports · 1 sources

Elias Hofele and other authors propose the TrySee modular closed-loop robot task planning pipeline, which combines general vision-language model (VLM) reasoning with domain-specific perception and existing robot components, requiring neither VLM fine-tuning nor end-to-end demonstrations. The system adopts reactive loop planning, adapting to different tasks through replaceable perception modules and motion planning, and adds domain-specific modular visual preprocessing to provide online grounding, with a dynamic preprocessing selection mechanism to improve scalability. The authors state that the method was evaluated on planning tasks across 5 substantially different domains, including a real robot disassembly task; field tests in brownfield environments show that the approach can integrate VLM planning into specialized robot systems when domain-specific training data is limited.