Self-Evolving Code Agents: From Digital Programs to Intelligence in the Physical World

Meta description: HexaAnything uses Physical Coding to lift RoboCasa365 Composite-Unseen success from 34.3% to 38.3% and overall success from 56.6% to 60.8% over XR-1 VLA.

Tags: Physical Coding, HexaAnything, RoboCasa365, XR-1 VLA, HexModel

Featured image: Abstract code grid with linked nodes and robot-like paths, no people, no logos, no text.

Physical Coding is the organizing idea behind HexaAnything, a robot learning interface that represents task state and execution as code. The latest reported evaluation gives it a measurable lift on RoboCasa365: Composite-Unseen success rises from 34.3% to 38.3%, and overall success rises from 56.6% to 60.8%, compared with the native XR-1 VLA. That shift matters because it shows a code interface can add structure to embodied task execution, not just produce more fluent language or longer action sequences.

Background: Why Code Became a Robot Interface

Robot systems that use vision-language-action, or VLA, models often convert perception and language instructions into motor behavior in one learned pipeline. The weakness is hidden state: a robot may know what to do next without carrying an explicit, inspectable record of what has changed, what remains blocked, and whether a plan is still valid.

Physical Coding addresses that gap by treating task state and execution as code. In this framing, Code as World records objects, relations, constraints, and progress. Code as Policy organizes actions, verification, and recovery. Together, they turn a robot task into something closer to a program that can be checked, revised, and resumed, rather than a purely reactive sequence of commands.

Development: Harness, World State, and Policy Loop

HexaAnything implements this interface through a Harness connecting models, tools, and external evaluation. The Harness creates a shared path for data collection and feedback, linking model behavior, tool results, and evaluation outcomes. In this setup, the world state can be checked, the policy can be revised, and the next data cycle can be shaped by what failed or succeeded.

This loop also shapes the training signal. A 27B HexModel trained on Harness-collected data improves over its base model on every split. That result places the Harness-collected data inside the reported improvement path, not as a separate side activity. The evidence is strongest where the evaluation compares the model with its base on each split.

Latest Progress: RoboCasa365 Gains and Broader Evidence

On RoboCasa365, the comparison against the native XR-1 VLA gives the clearest quantitative signal. Composite-Unseen success moves from 34.3% to 38.3%, and overall success moves from 56.6% to 60.8%. Both splits point in the same direction, which strengthens the case that the code interface is helping beyond a single narrow test case.

The evaluations also cover tool revision, simulated scientific experiments, and real-robot execution. Those areas matter because they test different failure modes: a revised tool changes the environment or interface assumptions, a simulated experiment demands structured procedure tracking, and a real robot brings sensor noise, timing limits, and physical consequences into the loop. Taken together, these results provide initial evidence for Harness, data, model, and tool improvement.

Broader co-evolution remains future work. The current stage is therefore not a complete closed loop between world representation, policy learning, and hardware execution, but a measured first pass. The next evidence will likely come from whether the same interface can keep improving as tools, environments, and model components change together, rather than whether it can win one more static benchmark.