Vision-language-action (VLA) and world-action (WAM) models map observations and instructions directly to robot actions, but that directness ties a policy tightly to its training distribution. According to a new paper posted on Hugging Face, even minor changes in layout or viewpoint can cause failure, and instructions generalize poorly.
The authors trace the problem to representation: task requirements, conditions, progress, and failure recovery are implicitly encoded in action sequences, making them hard to inspect or revise. They point to digital coding agents as a precedent, where LLMs call tools, verify results, and revise from feedback as executable code.
Their proposal, Physical Coding, represents task state and execution as code. Code as World records objects, relations, constraints, and progress, while Code as Policy organizes planning, verification, recovery, and execution. The system, HexaAnything, calls perception, planning, and control tools — including VLA/WAM policies — and makes in-the-loop decisions from external feedback.
On RoboCasa365, HexaAnything raises Composite-Unseen success from 34.3% to 38.3% and overall success from 56.6% to 60.8% over the native XR-1 VLA. A 27B HexModel trained on Harness-collected data improves over its base model on every split, which the authors read as evidence that code traces internalize physical execution.
Evaluations also cover tool revision, simulated scientific experiments, and real-robot execution. On PhyBench and a dual-arm AgileX robot, the agent autonomously completes physics experiments and most tabletop tasks, often faster than published results. The authors report initial evidence of data, model, and tool self-evolution, while broader co-evolution — including weight internalization and autonomous redesign of architectures, languages, representations, and tasks — remains future work.