SHAPER framework lets robots self-evolve skills without retraining models

A new arXiv paper from Peidong Wang and colleagues proposes SHAPER (Self-Harness Adaptation for Embodied Robots), a framework that lets embodied agents evolve their own behavior without any weight updates. The key insight: modern agents rely not just on frozen foundation models, but on the external skills, context prompts, action interfaces, and execution harness that surround them. SHAPER treats these non-parametric components as evolvable assets, improving them through rollouts in the target environment. The same frozen model acts as both planner and optimizer, refining its own tools and context code without gradient-based training.

In experiments across VLABench and ESI-Bench—which cover different low-level action interfaces—SHAPER was compared against pure execution, supervised fine-tuning, and test-time-scaling baselines like verifier-free selection and voting. The results show skill-and-harness optimization is a practical route to self-evolution when model training is time-consuming, costly, or impossible. This approach could make embodied agents far more adaptable in real-world settings where data, reward signals, and retraining runs are scarce—turning every deployment into a chance to improve.