Agentic Robot targets error build-up in long-horizon manipulation. It coordinates three parts through a protocol the authors call Standardized Action Procedure (SAP), modelled on standard operating procedures in human organisations: a large reasoning model splits the instruction into subgoals, a vision-language-action model executes each subgoal from live images, and a temporal verifier decides when a step is done and triggers recovery when it is not. The loop checks itself without outside supervision.
Findings
- LIBERO average success 79.6%, which the authors report as the best published.
- On long-horizon tasks it beats SpatialVLA by 6.1% and OpenVLA by 7.4%.
- The public code uses OpenVLA as the executor, DeepSeek-V3 as the planner and Qwen2.5-VL as the verifier.
Limits
- Abstract and repository README only.
- Evaluated on LIBERO in simulation, going by the abstract.
Relevance
The verifier is another model checking the executor, the same pattern as the VLM observer in RoboTwin 2.0 and the cross-check in GoGet. Open question for failure-recovery and runtime-monitoring: what happens when verifier and executor share a blind spot because both are VLMs trained on similar data?
Links
- arXiv · code
- Subfield: Robot learning, Safety and verification