Peidong Wang, Zhiming Ma, Ying Chang, Xufang Luo, Xiaocui Yang, Shi Feng, Yuqing Yang, Dongsheng Li
4 min
Embodied AI agent performance is determined not only by the underlying foundation model weights, but also by model-external components such as reusable skills, context construction, action interfaces, and execution harnesses. While supervised fine-tuning and reinforcement learning can adapt agents to new environments, they require expensive model updates, task-specific demonstrations, and additional reward data. Alternatively, train-free code-centric methods often rely on programmable robot APIs that may be unavailable in fixed-interface settings. This paper investigates whether an embodied agent can be effectively adapted in a train-free manner by optimizing its external procedural guidance and context construction around a frozen planner and executor.
The authors propose SHAPER (Self-Harness Evolution and Procedural Refinement), a train-free embodied adaptation framework that keeps all neural model parameters frozen while evolving a textual skill and a context-code harness. SHAPER factorizes the agent system into a frozen Vision-Language Model planner, a frozen executor (such as a Vision-Language-Action actor or an environment API wrapper), an optimizable textual skill, and an optimizable context-code harness. The adaptation process relies on rollout-guided textual diagnosis, where local round-level judgments and episode-level metadata are aggregated into a textual gradient. A frozen foundation model acts in an optimizer role—using a two-stage schedule to first refine the textual skill and subsequently update the context-code harness—while validating candidates via sandboxed evaluation and beam search.
The framework is evaluated on VLABench across held-out splits covering semantic understanding and common sense, and on ESI-Bench for embodied spatial intelligence. On VLABench, the Seed Agent achieves a success rate of 28.25%, outperforming direct VLA execution (23.25%) and same-data supervised fine-tuning (24.00%). Full skill-harness evolution further increases performance to 34.50%, outperforming test-time-scaling baselines such as verifier-free selection and trajectory voting. Furthermore, SHAPER exhibits strong generalization under distribution shifts, with larger performance gains when transferring to unseen target categories or unfamiliar task forms.
SHAPER demonstrates that non-parametric optimization of external skills and execution harnesses is a practical, effective alternative to parameter fine-tuning for embodied agents. By leveraging the same foundation model as both planner and optimizer, the framework enables autonomous self-evolution without requiring gradient updates, making it particularly useful when model training is expensive or interaction data are scarce.
Embodied agents are increasingly built as systems around foundation models, where performance depends not only on model weights but also on the skills, context, action interfaces, and execution harness surrounding the model. While supervised fine-tuning and reinforcement learning can adapt agents to new environments, they require additional data, rewards, and training runs; meanwhile, many train-free code-centric approaches rely on programmable robot APIs that may be unavailable in fixed-interface settings. We propose SHAPER, a self-evolving framework for train-free embodied adaptation that keeps model parameters frozen and improves the non-parametric agent system by evolving reusable skills and a context-code harness through target-environment rollouts. In SHAPER, the same frozen model can serve as both planner and optimizer, refining its external skills and context-code harness without parameter updates. We evaluate SHAPER on VLABench and ESI-Bench, covering embodied agents with different low-level action interfaces, and compare against pure execution, supervised fine-tuning, and test-time-scaling baselines such as verifier-free selection and voting. Our results suggest that skill-and-harness optimization is a practical route to self-evolving embodied agents when model training is expensive, unavailable, or undesirable.
Sam: But what are the limits? If the model itself isn't a strong reasoner, can better instructions really compensate for that?
Alex: That's the critical constraint the paper is upfront about. The whole system depends on the model being able to accurately diagnose its own failures. If it can't identify what went wrong, the written diagnosis will be vague or misleading, and the optimizer has nothing useful to work with. The process stalls.
Sam: It's like giving a detailed repair manual to someone who can't yet tell the difference between the parts. The manual can only help if the person reading it has enough baseline understanding to apply it.
Alex: That's a fair characterization. The paper frames SHAPER as a practical tool for situations where model retraining simply isn't available—not as a replacement for having a capable model to begin with.
Sam: So it's about squeezing more performance out of what already exists, rather than building something new from scratch.
Alex: That's the core claim. By letting the agent organize its own experience—diagnosing failures, rewriting its instructions, refining what it pays attention to—SHAPER turns a static system into one that can meaningfully adapt to new and unfamiliar environments, without anyone having to retrain it from the ground up. It's a measured but practical step forward in how we think about deploying robots in the real world.
Sam: That's a useful framing. Thanks for walking through it.
Alex: Thanks for listening to ResearchPod.