From VLM Planners to Robotic Agents: Self-Evolving Harnesses for History-Dependent Manipulation
- Published
- Source
- arXiv
- Paper number
- 1360
- Field
- Robotics
- arXiv ID
- 2610.29560
Key points
- SERA wraps the high-level VLM planner in an editable harness that maintains and updates task context using observations and execution feedback.
- Execution-driven self-evolution proposes revisions to planning guidance, assessment criteria, and control rules through environment rollouts and evaluates them with a verifier before harness selection.
- Self-evolution proceeded without constructing an additional supervised fine-tuning dataset for the high-level VLM.
- On the 16 RoboMME tasks, SERA with autonomous selection achieved 51.38% overall success, 7.00 percentage points higher than the 44.38% without self-evolution.
- SERA scored 69.50 on Reference, 57.00 on Permanence, 49.00 on Counting, and 30.00 on Imitation, obtaining the highest scores on Reference and Permanence.
- The authors note that VLM reasoning introduces planning latency into the execution loop, which accumulates during long-horizon execution as a limitation.
Paper links
External research summaries. These are not HDATF publications or measured product results.