Self-Evolving Coding Agents: From Digital Programs to Physical-World Intelligence
Hongcheng Gao, Jingjing Zhou, Zelin Zheng, Shijia Ge, Jay Zhu, Yazhe Wang, Jianshu Zeng, Xuan Shangguan, Di Wu, Lingyu He, Zhiqi Jia, Sihang Wu, Xiao He
cs.RO
2026-09-28
Robot task state and execution written as code: the same VLA, called as a tool, gains up to 31 points on long-horizon tasks; the traces train a 27B planner that beats GPT-5.6 in the same harness.
VLA models map an image plus an instruction straight to a chunk of robot actions. The paper opens with an uncomfortable measurement: a QwenGR00T policy trained on LIBERO with the instruction masked out succeeds in 92.3% of trials, against 96.2% for the instruction-conditioned version, with no suite gap larger than 6 points. When the scene determines the task, the policy fits scenes to trajectories; it is not following instructions. Released VLAs lose more than half their success under small viewpoint or initial-state changes, approach zero when the object layout is perturbed, and neither more data nor more capacity reliably fixes this.
The structural issue is the interface. A long-horizon task needs tracking of objects, constraints, unfinished subgoals, and recovery decisions, and none of that fits in an action chunk plus a stop signal. If the bread is in the oven but the skewer is still on the plate, a VLA closes the door anyway, and nothing in the interface can notice or repair it.
Physical Coding gives physical tasks the medium software coding agents already work in: executable code, shared by observation, action, and feedback.
Two design choices follow from the data loop rather than taste: because traces train the next model, the verifier cannot be the model that proposes actions, or it confirms its own errors; and because edits are inherited by the next version, admission has to be strict.
On RoboCasa365 with the action model fixed to XR-1, both coding agents driven by GPT-5.6-Sol, 50 seeds per task:
| Setup | Atomic-Seen | Composite-Seen | Composite-Unseen | Overall |
| XR-1 native | 78.0% | 54.8% | 34.3% | 56.6% |
| HexaAnything | 80.9% | 61.5% | 38.3% | 61.1% |
| Codex | 81.7% | 59.8% | 34.1% | 59.5% |
Codex does not beat bare XR-1 on Composite-Unseen, so the gain comes from state predicates and recovery branches, not from code generation alone. On three Composite-Unseen tasks at 100 seeds each: LoadKebabSandwich 17%→48%, PortionHotDogs 22%→37%, WaffleReheat 53%→70%. Doubling the VLA action budget moved native PortionHotDogs from 22/100 to 21/100, which rules out budget as the explanation.
For model evolution, 1,010 harness traces plus 9.3K VQA examples (38.8% of 178.6M training tokens) fine-tune Qwen3.8-27B into HexaModel v0.1. Placed back in the same harness, it beats its base on every split: Composite-Unseen 37.3%→39.5%, overall 60.5%→61.7%, 0.6 points above GPT-5.6-Sol (61.1%) in the same setup.
Tool evolution on RoboDojo, with task and evaluator frozen and only tool code revised: Fold cloth 0%→80%, Pour vase 40%→100%, Press by number 0%→100%, weights untouched. The path is not monotonic; the first Pour vase revision dropped it to 0% before the second reached 100%.
PhyBench, the paper's own Isaac Sim laboratory, covers three experiments: spring constant, gravitational acceleration, and coupled-oscillator frequencies, with success thresholds of 15%, 3%, and 10% relative error. With Opus 5.5 as the planner, HexaAnything lands at 1.3%, 1.0%, and 1.7% mean error with 10/10 valid runs on all three; with Qwen3.8-Max, only 5/10, 3/10, and 3/10 runs are valid. On a dual-arm AgileX robot, the abstract reports the agent completing physics experiments and most tabletop tasks autonomously, often faster than published results.
The 92.3%-versus-96.2% measurement matters on its own: high LIBERO-style scores do not show that a policy follows instructions, and such benchmarks deserve a discount.
The practical takeaway is the harness result. Without retraining or swapping the VLA, wrapping it in a code interface with state predicates and recovery branches lifts the hardest long-horizon case from 17% to 48%. Teams with existing VLA assets can treat this as a cheap incremental layer. The HexaModel numbers are the first evidence that verified execution traces are worth training on: an open 27B planner passing GPT-5.6-Sol inside the same harness, if only by fractions of a point.
On magnitude, cold water: model-level gains are incremental, and the experiments read as a roadmap with evidence attached; half the value sits in the framework and the measurements.
The authors scope the claims themselves: of the four evolution targets, the experiments cover the Harness, tool revision, and one data-to-model update; coordinated evolution of model, representation, environment, and embodiment is explicitly future work. HexaModel trains for one epoch and gains 0.2 to 2.2 points; whether the gain scales with more data is untested.
The Fold cloth evaluation carries an asterisk: a human operator picked the pixels, so the numbers measure the tool and interface rather than autonomous perception; the authors re-ran the final version without privileged information (3/5 development seeds, 4/5 held-out seeds), and seed 3 fails in every version. PhyBench is a new benchmark with no external systems to compare against. The compute cost of the admission pipeline (independent evaluators, regression tests, canary rollouts) is not reported, and on real hardware that overhead may not be small.