HiPhys-RSI: 2026 BEHAVIOR-1K Challenge Champion

Abstract

Scaling through executable structure

Can embodied intelligence scale through the construction and refinement of executable structure, rather than only through data, parameters and compute? Robotics already has strong tools: geometric reasoning, task and motion planning, SLAM, perception and feedback control. The bottleneck is the engineering needed to compose and adapt them for every new task. We present HiPhys-RSI (hierarchical physical recursive self-improvement), in which a frozen VLM coding agent automates that work. It examines execution feedback and videos, diagnoses failures and revises the code. Useful revisions are kept as reusable procedures and tools. The coding agent and its VLM backbone are never trained; all improvement is stored in executable procedures, tool compositions and structured knowledge. On the 2026 BEHAVIOR Challenge, the system ranks first across all 100 BEHAVIOR-1K household tasks, with a verified Q-score of 0.683 and a full-task success rate of 48.2% on the hidden test instances.

Official leaderboard

HiPhys-RSI leads the field

2026 BEHAVIOR Challenge, 100 tasks, hidden test instances, official evaluator.

View the official leaderboard on Hugging Face ↗

Top 6 verified entries

No other entry, verified or self-reported, scores higher; the best self-reported Q is 0.514. Source ↗

Method

A frozen agent that writes its way to better robots

Improvement lives in code and structured knowledge, not in model weights, so it carries over to the next task.

Experience the task

The agent reads the task specification and demonstrations, then controls the robot directly on a few training instances to learn which tools are needed and which observable conditions mean success.

Design the pipeline

Perception, navigation, planning, IK and control modules are composed into one autonomous program with state tracking, progress checks and conditional branches.

Analyze rollouts

Execution logs and VLM-based video analysis show where the physical outcome diverged from the plan, even when every tool call “succeeded”.

Refine iteratively

Targeted revisions change tool choice, parameters, execution order or recovery logic. The loop continues until the score stops improving.

Cross-task reuse

Finished pipelines become callable tools for new tasks, giving hierarchical composition: pipelines call reusable procedures, which call low-level skills.

Routing with policy cards

Each pipeline and VLA has a card recording its capabilities, historical success and known corner cases. The agent routes every task to the best-suited component.

Structure ⇄ data co-evolution

Trajectories from the pipelines fine-tune VLAs, and the improved policies return to the tool library, linking scaling by structure with scaling by data.

Results

Coding agents unlock the scaling of structured knowledge

Compared with four VLA variants built on the 2025 first-place π_0.5-based solution: meta-training on all 100 tasks for one or two epochs, per-task post-training, and 3D-fused per-task post-training.

Full category coverage, fewer manipulation failures

Each radial bar is one task, grouped into eight categories from mobile retrieval to object transformation. More VLA training raises individual scores but leaves whole categories empty; HiPhys-RSI fills the circle.

Rapid early gains come from composing capabilities; later gains come as RSI discovers new strategies. Most tasks converge quickly, and a long tail needs real exploration.

Placement, articulation and object-loss failures of the VLA baseline largely disappear. What remains is dominated by grasping and by execution-budget limits such as timeouts.

General placement and tool use need the most iterations; device operation needs the fewest.

Refinement iterations to
~76/100 on dev instances

~140

Median iterations
per task

25.5

Most-used
skill families

Task categories
covered

8/ 8

The plateau is not a ceiling

After the initial saturation, later gains come from strategies that appear nowhere in the demonstrations. They change how a task is approached, not just how it is tuned.

Failures shift

Compared with the VLA baseline, placement, articulation and object-loss failures largely disappear. The remaining failures are dominated by grasping and by execution-budget limits such as timeouts.

It generalizes

Pipelines were developed without access to the hidden test instances. Most tasks keep comparable scores on the hidden evaluation; the largest drops occur in search-heavy tasks with occluded objects.

Evaluation integrity

Benchmark Hacking or Reward-shaping?

A self-improving agent optimizing against a benchmark invites the question. These are the controls we used, and the gray-zone behaviours we found and report ourselves.

Hidden test never seen

Each pipeline was built and refined on a fixed set of 10 training instances per task. The hidden test instances stayed inaccessible throughout development.

Official evaluator, same budget

Scores come from the official evaluator with matched task-execution budgets across methods, so comparisons with VLA baselines are like for like.

Scored by the organizers

The 0.683 Q-score and 48.2% success rate were produced by the organizers on the hidden test instances and appear on the verified leaderboard.

Frozen agent, readable code

The coding agent and its VLM backbone are never trained. Pipeline improvements are stored as executable code, so every behaviour can be read and inspected.

It transfers to unseen instances

Most tasks score somewhat lower on the hidden evaluation than on development instances, but most of these drops stay below 0.2 Q, and several tasks hold or improve. The largest drops come from search-heavy tasks, where objects hidden in occluded places use up the time budget. Overall, the results indicate generalization across instances.

Where the line blurs

Under strong optimization, success predicates and partial rewards can stop tracking what a task intends. We looked for this and report two cases from our own analysis.

Instance 310 · 6×

Task 58 · stacking woodClear shortcut

The agent formed three separate pairs of logs instead of one woodpile. This satisfies the evaluator's condition that each log touches another, but not the intended stacking goal.

Instance 305 · 5×

Task 48 · canning foodAmbiguous

The agent diced food inside the refrigerator, earning partial credit while the later transfer-to-bowl goals stayed unmet. The evaluator correctly records incomplete progress, and cutting in place can be a valid intermediate step, but it may also reflect a pull toward easy partial credit.

Instance 302 · 4×

Not hacking: departing from demosLegitimate

Grasping a pan by its rim or tipping a sprayer before grasping (see the case studies above) differs from the demonstrations but achieves the task as intended. Deviation from demonstrations alone does not establish hacking.

Stronger systems expose gaps in task specifications; shortcuts we discover are reported so that evaluation criteria can evolve alongside the systems optimized against them.

Scaling by Structure

What forms can scaling take in embodied AI?