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

cs.ROSep 28, 2026

Self-Evolving Coding Agents: From Digital Programs to Physical-World Intelligence

Authors: Hongcheng Gao, Jingjing Zhou, Zelin Zheng, Shijia Ge, Jay Zhu, Yazhe Wang, Jianshu Zeng, Xuan Shangguan, +5 more

Abstract

Vision-language-action (VLA) and world-action (WAM) models map observations and instructions directly to robot actions. This directness ties a policy to training: minor layout or viewpoint changes cause failure, and instructions generalize poorly. The root cause lies in representation: task requirements, conditions, progress, and failure recovery are implicitly encoded in action sequences, making them difficult to inspect or revise. Digital coding agents offer a precedent: LLMs call tools, verify results, and revise from feedback as executable code. The same working pattern of explicit state, manageable execution, and revisable procedures underlies generalization and long-horizon execution in the physical world, letting physical experience return as reusable programs, memory, or evidence. We propose Physical Coding, representing task state and execution as code. Code as World records objects, relations, constraints, and progress; Code as Policy organizes planning, verification, recovery, and execution. We build HexaAnything, which calls perception, planning, and control tools, including VLA/WAM policies, and makes in-the-loop decisions from external feedback. Verified traces become data and memory, enabling evolution from tools and Harness to model weights, architectures, and ultimately hardware and task design. On RoboCasa365, HexaAnything improves Composite-Unseen and overall success over XR-1 VLA, and its Harness-trained HexaModel beats the base on every split, indicating code traces internalize physical execution. On PhyBench and a dual-arm AgileX robot, the agent autonomously completes physics experiments and most tabletop tasks, often faster than published results. We observe data, model, and tool self-evolution; future work targets weight internalization, autonomous redesign of architectures, languages, representations, and tasks, and deployment in manufacturing and science.

Figures & tables

Figure 1 : Physical Coding brings AI into the physical world through an explicit executable interface. Code as World represents task-relevant objects, relations, observations, constraints, and progress predicates, while Code as Policy organizes planning, tool calls, action execution, verification, and recovery. Together they form the core interface through which the agent interacts with physical environments and accumulates reusable artifacts for continual improvement.

Panel of Figure 2 — Figure 2 : Removing the instruction barely changes behavior on LIBERO. (a) Masking the instruction during training lowers success by at most 6.0 percentage points; both policies are QwenGR00T in StarVLA (Qwen3-VL-4B backbone, GR00T-style action head) trained jointly on the LIBERO training data, with 500 trials per suite. (b) Given the instruction or an empty prompt, the same VLA picks up the alphabet soup and places it in the basket. The VLA is a LIBERO-finetuned $\pi_{0.5}$ [ 64 ] .

Figure 3 : Physical Coding: coupling world state, policy execution, and evidence. Physical observations and tool outputs are translated into Code as World , which represents task-relevant objects, relations, measurements, constraints, and progress predicates. Code as Policy uses this representation to organize planning, tool calls, action execution, verification, and recovery. Execution produces new observations that update the world representation, while verification against fresh evidence determines whether to continue, re-observe, recover, or stop.

Figure 4 : Physical self-evolution. Code as World and Code as Policy provide a shared, evolvable interface across four targets: the Harness, data and environments, model and algorithms, and embodiment and compute. Execution evidence guides candidate edits, independent evaluation determines acceptance or rollback, and validated records support subsequent learning and refinement.

Figure 5 : Task inputs guide a physical-world model that produces world and policy code for the Harness’s Observer–Plan–Act–Verify–Recover loop. Actions interact with simulated or physical environments, while verified traces feed training data and subsequent model, Harness, and environment improvement.

Split	XR-1 (native)	HexaAnything	Codex
Atomic-Seen	78.0%	80.9%	81.7%
Composite-Seen	54.8%	61.5%	59.8%
Composite-Unseen	34.3%	38.3%	34.1%
Overall	56.6%	61.1 %	59.5%

Table 1: Harness comparison on RoboCasa365. XR-1 runs natively or is called as a tool by a coding agent; both coding agents are driven by GPT-5.6-Sol. Entries are success rates. Each task is evaluated on 50 seeds (18 Atomic-Seen, 16 Composite-Seen, and 16 Composite-Unseen tasks), and Overall pools all trials.

Task	XR-1 (native)	HexaAnything
LoadKebabSandwich	17.0%	48.0%
PortionHotDogs	22.0%	37.0%
WaffleReheat	53.0%	70.0%

Table 2: RoboCasa365 Composite-Unseen case study. Entries are success rates over 100 seeds. HexaAnything, driven by GPT-5.6-Sol, calls the same XR-1 as a tool inside an editable workflow.

Panel of Figure 6 — Figure 6 : LoadKebabSandwich from the same initial state (seed 3) . Native XR-1 puts the bread in the oven and closes the door with the kebab still on the plate (red circle); the kebab can no longer be placed. HexaAnything checks after each XR-1 call which items are inside and allows the door to close only when both are (green circle).

Stream	Content	Records	Tokens
Agent traces	subgoals, tool calls, verifier outcomes, recovery decisions	1,010	42.4M
VQA with CoT	RoboCasa365 state and task-completion judgments; general embodied VQA	12,075	41.7M
Scene captions	RoboCasa365 scene descriptions	6,764	2.0M
General data	math, code, and instruction following	24,491	92.5M
Total		44,340	178.6M

Table 3: Training data of HexaModel v0.1, trained for one epoch. The agent traces and the 9.3K RoboCasa365 VQA examples are returned by the Harness.

Planner	Atomic-Seen	Composite-Seen	Composite-Unseen	Overall
XR-1 (native, no Harness)	78.0%	54.8%	34.3%	56.6%
Qwen3.8-27B (base)	80.8%	61.0%	37.3%	60.5%
GPT-5.6-Sol	80.9%	61.5%	38.3%	61.1%
HexaModel v0.1	81.0%	62.3%	39.5%	61.7%

Table 4: Model evolution on RoboCasa365. Each row is the planner inside HexaAnything, except XR-1 (native), which runs the VLA without the Harness. Entries are success rates. Each task is evaluated on 50 seeds (18 Atomic-Seen, 16 Composite-Seen, and 16 Composite-Unseen tasks), and Overall pools all trials.

Panel of Figure 7 — Figure 7 : Success rates of Table 4 by split on RoboCasa365. XR-1 (native) runs the VLA without the Harness; the other three bars are the planner inside HexaAnything with XR-1 as the action tool. Each task is evaluated on 50 seeds (18 Atomic-Seen, 16 Composite-Seen, and 16 Composite-Unseen tasks).

Revision round	Fold cloth	Pour vase	Press by number
Round 0	0.0%	40.0%	0.0%
Round 1	60.0%	0.0%	0.0%
Round 2	80.0%	100.0%	100.0%

Table 5: Tool self-evolution on three RoboDojo tasks. Each cell is the task success rate after the corresponding revision round; the task definition and the evaluator are the same in every round.

Panel of Figure 8 — Figure 8 : Tool self-evolution on RoboDojo Fold cloth, development record. Each point is one version of the pick-and-place tool, placed at the development iteration in which it landed and replayed on seeds 0–4 with one fixed plan per seed; for three of the five seeds that plan was chosen with access to diagnostic garment keypoints, so the re-evaluation of the final version without privileged information (3/5 development seeds, 4/5 held-out seeds) is reported in the text. v0 rejects the garment as too wide; v1–v2 add the pinch mode for thin layers; v3 synchronizes the two arms; v4 tilts the wrists apart; v5 releases vertically; v6 raises the drop point over folded layers; v7 separates the hover clearance at the drop.

Figure 9 : The three PhyBench experiments and the robot operations they require. (A) Hooke’s law: place weights on the spring-supported tray and read the ruler. (B) Simple pendulum: push a bob with a magnetic probe, withdraw, and time complete cycles. (C) Coupled oscillators: displace a slider, pull the release bar, and record both sliders. Frames are from simulation runs and show one feasible procedure; the agent designs its own.

Figure 10 : A Hooke’s-law run with HexaAnything and Opus 5.5. One of the ten runs in Table 6 ; the plan (top) precedes any action. (a) One loading cycle, repeated for each of the four weights: segment, grasp, place on the tray, move clear of the ruler, and read the pointer between the 22 and 20 cm ticks. (b) Readings at the five load levels, each with its source observation (obs $n$ ), and the weighted linear fit: $k=24.82\pm 0.40$ N/m against a reference of 25.16 N/m (error 1.4%). Masks and slot grids are tool outputs; circles, dots, and lines are added.

Harness + Model	Hooke’s law	Simple pendulum	Coupled oscillators
Harness + Model	$k$ error $\downarrow$	$g$ error $\downarrow$	$\omega_{1,2}$ error $\downarrow$
HexaAnything + GPT-6-Astra	2.3% (10/10)	1.2% (10/10)	0.8% (10/10)
HexaAnything + Opus 5.5	1.3% (10/10)	1.0% (10/10)	1.7% (10/10)
HexaAnything + GPT-5.6-Sol	4.8% (9/10)	1.9% (8/10)	1.2% (10/10)
HexaAnything + Qwen3.8-Max	12.4% (5/10)	2.4% (3/10)	3.7% (3/10)

Table 6: Scientific experiment execution with HexaAnything in simulation. Entries give mean relative error (%; lower is better) over valid runs and the valid-run count

n/N

Figure 11 : Initial scenes of five of the seven real-robot tasks, seen from the front scene camera: unscrew bottle cap, hot-dog serving, toss blocks into bowl, fold clothes, and pour blocks. Tic-tac-toe and block into bowl are not pictured.

Task	Success	Progress	Output tokens	Time (min)	Published reference (same model)
Unscrew bottle cap	3/3	100%	1.52k	5.1	GPT-Policy R3: 3/3, 17.9 min
Tic-tac-toe	3/3	100%	3.26k	5.0	GPT-Policy R8: 3/3, 13.6 min
Block into bowl	3/3	100%	443	1.0	Robocurve: 19/20, 2.1k tokens, 2.5 min
Hot-dog serving	3/3	100%	2.40k	7.9	–
Pour blocks	3/3	100%	2.29k	3.1	–
Toss blocks into bowl	1/3	77.8%	4.69k	8.1	–

Table 7: Seven tabletop tasks on the AgileX dual-arm platform, each run three times with GPT-6-Astra as the execution agent calling URAI tools. Progress is the fraction of the task completed (blocks in the bowl out of six for toss blocks; the operator’s completion score for fold clothes). Tokens are the execution agent’s output tokens per episode, reasoning included; time is wall-clock from task release to a confirmed outcome and includes model latency, robot motion, and, in tic-tac-toe, the human’s moves. Published references use the same model on other hardware (GPT-Policy on its own robot; Robocurve on YAM arms with a 25% speed cap and at most 20 model calls) and are not same-robot measurements. The first unscrew-bottle-cap trial excludes an initial diagnosis-and-repair phase.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Line of work and representatives	Primary evolving artifact	Persistence and typical fixed boundary	Feedback or evidence	Relation to this work
Task, scene, and curriculum generation: GenSim, RoboGen, SceneSmith, CurricuLLM [ 82 , 87 , 63 , 70 ]	Task specification $z$ , scene/environment $e$ , or curriculum	Generated tasks and scenes persist as training or evaluation assets; the downstream policy, learner, and evaluator are usually external.	Task solvability, policy performance, or curriculum progress.	Covers code as world and environment evolution; it does not by itself revise the executing policy and verifier together.
Data and trajectory synthesis: GenSim2, MimicGen, RoboTwin, HumanoidGen [ 30 , 56 , 59 , 36 ]	Data $D$ : demonstrations, trajectories, captions, or VQA	Data is retained for later training; the generator and downstream update rule are usually fixed for a run.	Dataset scale or quality and downstream policy success.	Motivates data evolution; our traces link training data to state predicates, recovery, and held-out validation.
Reward and objective search: Eureka, DrEureka, REvolve [ 55 , 54 , 27 ]	Reward code, domain randomization, or preference-derived objective	Candidate objectives are iterated within training; the policy and evaluation boundary typically remain specified externally.	Rollout return or success plus automated or human feedback.	Shows objective evolution, but a training reward is not an independent verifier $\phi$ .
Evaluation and verifier generation: RoboPlayground, AutoEval, SimFoundry [ 86 , 105 , 67 ]	Task specifications, resets, success predicates, and evaluator	Evaluation artifacts can be reused across policies; policy training is usually held fixed during comparison.	Terminal success, process evidence, reset reliability, or evaluator agreement.	Directly informs $\phi$ and provenance; our verifier is coupled to recovery and regression tests while remaining independent.
Programmatic policy synthesis: Code as Policies, RoboCodeX, RoboScript, RoboCoder, CaP-X [ 47 , 58 , 10 , 41 , 21 ]	Executable policy programs and skill compositions	Programs execute per task or are reused as skills; the base model, tool semantics, and evaluator are generally fixed.	Execution success, task completion, and code or skill validity.	Closest precedent for code as policy; our world program makes intermediate state and constraints explicit.
Tools, interfaces, and agent runtimes: ReAct, SWE-agent, OpenHands, Model Hardware Standard [ 98 , 95 , 85 , 3 ]	Tool schema, agent-computer/robot interface, or hardware driver	The interface structures observations and actions; model weights and task/evaluation protocols generally remain fixed.	Tool execution, repository tests, or interface-level correctness.	Supports modular boundaries and physical adapters, but interface standardization alone is not self-evolution.

Appendix

Table 8 : Mechanism-level comparison of related work. Rows group representative systems by the artifact whose adaptation is central. “Fixed boundary” summarizes the usual reported setup and is not a claim about every implementation.

Panel of Figure 12 — Figure 12 : PortionHotDogs, seed 25. Native XR-1 completes the right plate and puts bread on the left one, but does not return to the sausage left in the bowl; the left plate ends with bread only (red circle). HexaAnything detects the missing item and issues a grasp for the sausage in the bowl, which XR-1 then places.

Panel of Figure 13 — Figure 13 : PortionHotDogs, seed 33. Native XR-1 keeps grasping inside the bowl, and the light-blue plate stays empty (red circle). In the HexaAnything run, a bread falls onto the counter beside the bowl (red circle); the agent detects the failed placement and has XR-1 put that bread on the light-blue plate (green circle).

Panel of Figure 14 — Figure 14 : PortionHotDogs, seed 46. Native XR-1 portions both breads and then stalls with both sausages in the bowl (red circle). HexaAnything checks task progress and issues grasp-and-place commands for the sausages; each plate ends with one bread and one sausage (green circle).

Figure 15 : Simple-pendulum example run. The plan (top) is written before any action. (a) One excitation, shown for P2: the agent marks grasp points on the magnetic probe and picks it up, the wrist camera shows the attached probe, the agent marks a push point on the bob (circled), pushes and withdraws, and reads the ten-cycle time from the gate timer. The same steps are repeated for P1 and P3. (b) Gate records, each with the gate-timer record it was read from, and the zero-intercept fit of $T^{2}$ against $L$ over the three pendulums; the lower strip shows residuals.

Figure 16 : Coupled-oscillator example run. The plan (top) is written before any action. (a) Preparation of the second record: the agent segments the CH2 post and selects grasp points on it (gripper above the post), displaces the CH2 slider to $-23.9$ mm (wrist camera), marks grasp points on the release bar, and pulls it to start a 120 s two-channel recording. The first record used the same steps with CH1 at $+31.7$ mm. (b) Both records with the joint two-mode fit over 0–45 s, and the normal-mode frequencies of each record and of their mean, shown as deviations from the reference.

Record	Initial displacement	$\omega_{1}$ (rad/s)	$\omega_{2}$ (rad/s)
1	CH1 $+31.7$ mm	3.0139	3.3382
2	CH2 $-23.9$ mm	3.0141	3.3383
Estimate (mean of records)		3.0140	3.3382
Reference		3.0137	3.3376

Appendix

Table 9: Independent records and estimated normal-mode frequencies in the coupled-oscillator run.

Explore similar work

Sep 24, 2026cs.RO

HarnessPAI: An Evolving Harness for Physical AI

Physical AI aims to build embodied agents that perceive the world, understand and reason about it, and decide how to act. Yet the field has focused primarily on the last component: the action model that maps observations to low-level controls. The prevailing training recipe can erode the perceptual and reasoning capabilities needed for robust behavior, leaving even strong action models vulnerable to scene perturbations and long-horizon tasks. We introduce HarnessPAI, a model- and embodiment-agnostic Harness framework for Physical AI that treats code as the executable and evolvable interface that organizes the underlying action primitive. The framework separates two timescales: within a rollout, it executes open-loop at the program level, with a fixed program guiding and checking execution; across rollouts, it evolves closed-loop, using execution feedback to revise the program and distill failures into reusable skills. Across desktop robot arms, household robots, a robot vacuum, and a legged walking agent, HarnessPAI improves on both pure action models and code-as-policy baselines without retraining the underlying model: a 61.6-point gain over

π_{0.5}

on LIBERO-PRO and a 27.2-point gain over WorldDreamer on RoboCasa atomic tasks. Once a program is selected, rollout execution requires no online high-level LLM deliberation. Beyond execution, the converged program is also a cheap and reliable expert-data collector, and fine-tuning

π_{0.5}

on collected expert data lifts success rate on LIBERO-PRO by 38.8 points. Our results suggest that the frontier of Physical AI depends not only on stronger action models, but also on executable harnesses that integrate perception, task understanding and reasoning, and action execution into a unified, verifiable, and feedback-driven system. Website: https://darwin-agent.github.io/HarnessPAI

Xin Wang, Wenhao Wu, Menghao Zhang +20

Embodied Artificial Intelligence Agent Harness Embodied Agents Perception Pipeline Artificial Intelligence Systems Evolving Harness+1

Ask PDF

Jul 13, 2026cs.RO

From World Action Models to Embodied Brains: A Roadmap for Open-World Physical Intelligence

Artificial general intelligence ultimately requires agents that can reason and act in the physical world. Action models, vision-language-action policies, and world models have advanced this goal, while World Action Models (WAMs) are particularly promising because they connect candidate interventions with predicted consequences. However, progress remains fragmented: models use incompatible action spaces and prediction targets, datasets and tasks follow different conventions, and runtime systems expose limited interfaces for reuse and evaluation. We review the evolution toward WAMs and organize these limitations into three coupled gaps: model roles and representations, objectives and standardization, and system composition. Building on this analysis, we propose a co-evolution roadmap for physical intelligence centered on the \emph{embodied brain}, a long-term model target for integrating multimodal context, comparing candidate interventions, and issuing state-transition or capability requests rather than direct actuator commands. WAMs provide promising prototypes for its predictive functions, while a physical harness grounds model outputs through tools, controllers, verification, and trace logging. Shared contracts align heterogeneous models, data, tasks, and embodiments, and closed-loop post-training converts verified interaction into reusable experience. Together, these components define a modular physical-intelligence stack for adaptive and self-improving embodied agents.

Yuanzhi Liang, Xufeng Zhan, Haibin Huang +2

Embodied Artificial Intelligence World Models General Intelligence Embodiment Artificial Intelligence Systems Roadmap+1

Ask PDF

Aug 3, 2026cs.AI

Towards the Harness of Embodied Agents

The success of coding agents has established the harness as a paradigm: what an agent achieves depends not on the model alone, but on the infrastructure around it. We ask whether the same paradigm extends to embodied agents in the physical world. We present Thea, a harness in which an agentic loop orchestrates robot capabilities, each wrapped as a callable tool. It inherits the core components of coding agents, modified as the physical world requires. The world, however, withholds two abilities that software grants for free: reading the state of the world, and judging the outcome of an action. To bridge these gaps, Thea introduces Scene Graph as Context, a persistent, symbolic representation of the world, and Evaluation as Exit Codes, which detects when an action should terminate, judges whether it succeeded, and on failure diagnoses the cause. Together they close the loop between the agent and the physical world. Rich behaviors then emerge from the composition of tools, and the closed loop carries long-horizon tasks to completion in real environments.

Qi Wang, Tianyi Wang, Chengyang Li +6

Eastern Institute of Technology, Ningbo

Embodied Agents Agent Loop Robot Systems Long-Horizon Task Planning

Ask PDF