PaperlayerView exact revision on arXiv

Source-linked AI summary

Scaling Online Exploration for Robot Agents

Runhan Huang, Xiangcheng Zhang, Haoyu Chen, Yitong Li, Yen-Jen Wang, Bohan Lyu, Haotian Ye, Chuye Hong, Yixuan Wang, Chloe Su, Haoyu Xiong, Yangzhen Wu, Enze Xie, Travis Manderson, Stefano Ermon, Pieter Abbeel, Antonio Torralba, Dawn Song, Weirui Ye, Yilun Du

arXiv:2610.22496v1cs.RO

TL;DR

Robot policies trained on demonstrations remain bounded by imitation, while new robots and tasks typically require dedicated teleoperation data. RAX lets a coding agent improve through online exploration, compiling successful decision sequences into executable robot programs. Exploration improves performance across unseen environments, including a +37.4-point average advantage over top baselines in simulation.

  • Problem

    Demonstration-trained robot policies are bounded by their imitation distributions, and deploying on a new robot or task typically requires a dedicated teleoperation corpus.

    Source: §1 · p. 5

  • Method

    RAX uses a general-purpose coding agent to explore through robot interfaces and compile successful interactions into autonomous programs and reusable skills, without policy training.

    Source: §1 · p. 5

  • Results

    +37.4 points on average across five simulation benchmarks over state-of-the-art baselines, with performance improving as exploration scales and real-robot tasks learned without demonstrations.

    Source: §1 · p. 7

  • Takeaways & Limitations

    RAX represents accumulated experience as inspectable, verifiable programs and skills that can be composed and reused across episodes and tasks.

    Source: §1 · p. 7, §6 · p. 30

Abstract

from arXiv · show

Robot learning has advanced by scaling demonstration datasets and model capacity; however, deploying a new robot embodiment or task typically necessitates a dedicated teleoperation corpus. The resulting policies remain bounded by their imitation distributions, limiting generalization and continual self-improvement. We introduce RAX (Robot Agent eXploration), an open-source framework for robot agents centered on scaling exploration: enabling agents to improve by acquiring online experience and synthesizing executable programs. In RAX, a general-purpose coding agent generates robot control code, operating without action demonstration data, hand-engineered action primitives, or value-function guidance. Through online exploration, the agent identifies decision sequences that achieve task goals and compiles them into modular programs encapsulating the full perception-planning-control loop. Using RAX, we present a systematic multi-embodiment study on scaling online exploration across precision and long-horizon manipulation tasks using single-arm, bimanual, dexterous, mobile, and humanoid systems in simulation and on real robots. Generalization performance on unseen environments improves consistently as exploration scales. In simulation, RAX outperforms top baseline models by +37.4% on average. On real robots, RAX autonomously learns memory-dependent, articulated, high-precision, deformable, and long-horizon mobile manipulation tasks without human demonstrations, continually improving performance by composing verified skills from prior exploration sessions. As exploration scales, we observe emergent capabilities including active perception, error recovery, safety guardrails, and human-robot collaboration. We open-source RAX as a generalist robot-agent platform in which new robot embodiments are integrated through their APIs alone, without teleoperated data collection or policy training.

1 Introduction

RAX addresses the dependence of robot learning on task- and embodiment-specific demonstrations by letting a coding agent learn from its own online experience. Across simulation and real robots, exploration improves performance, while executable programs and reusable skills support adaptation and increasingly complex tasks.

1 Introduction

  • Demonstration-trained policies are bounded by their imitation distributions, and new robots or tasks typically require a dedicated teleoperation corpus.The introduction frames autonomous online experience and active exploration as a question for robot self-improvement and skill discovery.Source: §1 · p. 5
  • Rather than imitating actions, the agent programs decision-making from perception and tests and revises those programs through online execution.The programs integrate perception, planning, and control, and can be composed or updated as scenes and tasks become more complex.Source: §1 · p. 5
  • RAX lets a general-purpose coding agent explore through robot interfaces, then compile successful interactions into autonomous programs and reusable knowledge without policy training.The agent receives a natural-language instruction, at most a few goal images or visual demonstrations, and an exploration-episode budget; it requires neither teleoperation nor value-function guidance.Source: §1 · pp. 5, 7
  • Scaling exploration reveals 3D perception, pose-search planning, closed-loop visual and force feedback, and adaptive programs, alongside active perception, error recovery, safety checks, and human interaction.Examples include contact probing when depth is inaccurate and adding a workspace-clearance check after a cabinet collision.Source: §1 · pp. 5–6
  • Across household sessions, RAX records failures and successful strategies, then stores verified modular skills with code, logs, preconditions, and checkpoints for downstream composition.On a composite mobile-manipulation task with more than six sequential object manipulations, reusing stored skills yields substantial automation and efficiency gains.Source: §1 · pp. 6–7
  • +37.4 points on average over five simulation benchmarks: RAX outperforms state-of-the-art baselines while learning diverse manipulation tasks through autonomous exploration.The study also reports zero-demonstration real-robot results across high-precision, deformable, articulated, and long-horizon tasks.Source: §1 · p. 7
  • RAX releases its architecture, instructions, skills, integrations, and evaluation protocols as a platform for bringing up new robots through interfaces alone.The platform requires neither data collection nor policy training for a new robot.Source: §1 · p. 7

2 Related Work

Prior robot language-model methods commonly generate plans or programs with little opportunity to improve through experience. RAX instead treats the agent as a learner, positioning online exploration and accumulated experience as alternatives to demonstration scaling and as a route to broader embodiment coverage.

2 Related Work

  • Language-model planners and program generators often produce plans or programs zero-shot or from a few examples, with little opportunity to improve from experience.RAX instead revises perception and control programs based on observations gathered over multiple robot episodes.Source: §2 · pp. 7–8
  • Prior agentic robot-learning work has mainly focused on tabletop parallel-jaw manipulation, often using one platform, demonstrations, offline data, reconstructed simulators, or neural policy training.RAX studies exploration across single-arm and bimanual manipulators, dexterous hands, mobile manipulators, and legged humanoids, including systematic scaling with exploration, 3D perception, and accumulated memory.Source: §2 · p. 8
  • Robot-learning approaches scale demonstrations and model size, while generalist policies require data and fine-tuning for robots outside their training mixture.RAX studies capability growth from online experience gathered by the agent, without additional demonstrations or parameter updates.Source: §2 · p. 8
  • Prior language-agent memory systems retain skills or textual lessons, while RAX’s skills are executable control code that enters memory only after succeeding without supervision.RAX also draws on cognitive architectures that organize behavior around perception, memory, and deliberation.Source: §2 · pp. 8–9
  • RAX’s architecture connects the coding agent to hardware through perception and control interfaces, with persistent memory for context, notes, and verified skills.It compiles perception and planning into standalone Python programs that execute high-frequency loops on the robot; failures can return control to the agent for code refinement.Source: Figure 2 · p. 9

3 The RAX Cognitive Architecture

RAX builds task-specific robot skills through exploration, combining multimodal perception, 3D planning, executable programs, persistent memory, and parallel exploration. These components support continual skill reuse and faster exploration across tasks.

3 The RAX Cognitive Architecture

  • RAX’s cognitive architecture acquires task-specific perception, manipulation strategies, and executable skills from the agent’s own experience.It combines perception and control, autonomous robot programs, and memory for continual learning.Source: §3 · p. 9

3.1 Multimodal Perception

  • RAX builds 3D world models from depth and regulates contact with force feedback, with perception and object-model fitting emerging through exploration.For the air fryer, the agent models a sliding drawer and plans grasping, pulling, and releasing on that articulated model; the predicted basket rim differs from the post-opening measurement by 6.5 mm.Source: §3.1 · pp. 9–10, Figure 3 · p. 10

3.2 Planning and Control

  • Given its world model, the agent searches for feasible robot trajectories while accounting for posture, collisions, and joint limits.For the air fryer, it pulls in 3 cm steps and tilts the hand by 60° to place the banana.Source: §3.2 · p. 10

3.3 From Exploration to Autonomous Robot Program

  • RAX compiles explored decisions into modular programs that check preconditions, execute motions, and verify outcomes, then tests them unattended in new scenes.Exceptions let the agent intervene when open-world variation defeats perception or planning; programs can adapt to new layouts rather than replaying trajectories.Source: §3.3 · p. 11

3.4 Continual Learning with Memory

  • RAX retains session experience and verified skills across tasks, letting later sessions retrieve and compose successful programs instead of rediscovering behaviors.Practice refines subtask programs before they are committed to a skill library for long-horizon tasks.Source: §3.4 · pp. 11–12

3.5 Parallel Multi-Agent and Multi-Environment Exploration

  • More than 2× less wall-clock time is required to reach target performance when simulation exploration is parallelized across agents and environments.Agents evaluate distinct strategies concurrently rather than sequentially in one environment.Source: §3.5 · p. 12

4 The RAX System

RAX separates the agent’s frontend from each robot’s backend through a tool interface. Shared agent programs and interfaces support multiple embodiments while backend controllers handle low-level control and safeguards.

4 The RAX System

  • RAX uses one frontend across simulated and real embodiments, while each backend retains its robot infrastructure and safety enforcement.The agent acts through tools and cannot modify controllers or drivers; the backend handles low-level control and safeguards.Source: §4 · p. 12, Figure 4 · p. 13

4.1 Frontend–Backend Separation

  • A new robot can join RAX through a backend exposing its tools and documented interfaces, without teleoperation data collection or policy training.The frontend, memory organization, and evaluation protocol are shared across platforms.Source: §4.1 · pp. 13–14

4.2 Multi-Embodiment Control Interface

  • RAX combines backend high-rate tracking with agent decisions at the level of short action chunks and motion goals, while exposing raw control and model-based helpers.The agent can choose its abstraction level, from joint targets to planned free-space motions; Table 1 summarizes low-level control by platform.Source: §4.2 · p. 14, Table 1 · p. 15

5 Experiments

The experiments evaluate RAX across simulated and physical robot tasks, examining how exploration scales, how perception and parallelism affect performance, and how agent behavior changes. They also assess learning behaviors such as perception, error recovery, and collaboration, alongside reuse and adaptation of verified programs as tasks and environments change.

5.2.1 RoboDojo

  • 78.3% average across seven RoboDojo tasks, versus 24.9% for VPP2, with RAX exceeding every baseline on every task at its largest budget.Performance generally rises with exploration: pour by language reaches 100.0%, and make kong rises from 64.0% at 16 episodes to 96.0% at 32; domain-randomized tasks need larger budgets.Source: Table 2 · p. 17, §5.2.1 · p. 17

5.2.2 BEHAVIOR-1K

  • On BEHAVIOR-1K, cooperating agents explore six long-horizon household tasks for 128 episodes, receiving completion keyframes and post-run review reports before evaluation on ten public test instances.The review model identifies missed goal conditions and likely causes, suggests code changes, and reports partial scores with camera frames; no benchmark result is supplied here.Source: §5.2.2 · pp. 17–18

5.2.3 HumanoidArena

  • 89.8% average success on HumanoidArena, 7.9 points above WB-WAM and 29.8 above the strongest benchmark baseline; without action data, RAX ⋆ reaches 81.9%, matching WB-WAM.RAX succeeds in every OpenDoor and Boxing test episode; without action data, it remains within 5 points of RAX on Football, OpenDoor, and SitSofa, but trails by 10.0–16.6 points on dynamic, contact-rich tasks.Source: §5.2.3 · pp. 18–19, Table 4 · p. 20, Table 10 · p. 25

5.2.4 DexJoCo

  • 96.7% average success across eleven DexJoCo tasks at 32 exploration episodes; the best session per task reaches 99.5%, 45.4 points above the strongest benchmark baseline, without action demonstrations.RAX reaches 100% on Hanoi and iPad, and 96.0% on Assembly, where the best published result is 28.0%.Source: §5.2.4 · p. 20, Table 5 · p. 21

5.2.5 LIBERO and LIBERO-Pro

  • 98.5% average success on original LIBERO suites without action demonstrations; on position-perturbed layouts, RAX reaches 76.0% without adaptation and 84.4% after up to five episodes.With no perturbed-layout exploration, RAX exceeds every learned policy on every suite by 33.7 points over VPP2; after adaptation, it exceeds ASPIRE on every suite.Source: §5.2.5 · p. 21, Table 6 · p. 22

5.3.1 Parallelization; 5.3.2 Depth versus RGB-Only Perception

  • Parallel exploration reduces wall-clock time, while depth-based perception scores 17.3 points higher on average than RGB-only perception across matched RoboDojo task and budget pairs.With four agents, Claude Code scores 92.2 in 4.9 hours versus 63.3 in 13.7 hours for a single agent with a larger budget.Source: Table 7 · p. 22, §5.3.1 · p. 22, Table 8 · p. 23

5.3.3 Inference Time and Tokens

  • Compilation cuts inference overhead as exploration progresses: loading-the-car inference falls from 123 minutes and 254k tokens to 4 minutes and 4k tokens, while air-fryer program runs take 18–23 seconds.On make toast random, final agent-monitored successes take 6 seconds per episode; a compiled Bandu program assembles seven blocks with 10 seconds of program overhead.Source: Figure 7 · p. 24, §5.3.3 · pp. 24–25, Table 9 · p. 25

5.4.1 Active Perception

  • As exploration scales, RAX exhibits active perception by moving its cameras and hand to gather task-relevant measurements, rechecking before actions, and measuring outcomes afterward.Source: §5.4.1 · p. 25
  • RAX used active perception to calibrate robot geometry and surroundings, including force-based probing that corrected camera offsets and enabled a peg to enter its hole on the first probe.In simulation, it inferred Allegro fingertip offsets within 1.1 mm of the official model and measured the Hanoi disk height with an average 0.66 mm prediction error. On the real peg-in-hole task, it measured board and table heights and corrected wrist-camera positions using touch.Source: §5.4.1 · p. 26, Figure 8 · p. 26

5.4.2 Error Recovery and Safety

  • After physical mistakes, RAX encoded corrective rules into later programs, while simulated humanoid falls informed safer foot placements for crossing a bar.On the RB-Y1 it checked base moves against the swept area and stood before lifting a bucket; on the Franka it restricted arm postures near cameras and recalibrated a bumped camera.Source: §5.4.2 · pp. 27–28, Figure 9 · p. 27

5.4.3 Human–Robot Interaction

  • RAX used operator requests to secure or reposition objects and to test generalization, succeeding unattended across new fryer placements except one that exposed errors it then fixed.Across runs 6–10, the fryer moved by up to 7 cm and its handle rotated by up to 18°; the agent also refused a banana pick that risked hitting a camera and asked for it to be moved.Source: Figure 10 · p. 27, §5.4.3 · p. 28

5.5.1 Reusing Skills

  • On the RB-Y1, RAX composed stored, verified skills into a six-stage tidy-up task, refined affected programs, and added synthesized control code to its reusable repository.Skills were versioned with precondition checks; across three iterations, revisions ranged from small cabinet-measurement updates to deeper chair grasps and support for additional cup-stacking objects. Newly synthesized hand-reach code was abstracted and reused across chair and cup routines.Source: §5.5.1 · pp. 28–30, Figure 11 · p. 28

5.5.2 Rapid Adaptation of Programs

  • On the Franka, RAX adapted programs by preserving valid modules while adding task-specific perception and paper-handling routines, completing the revised cube tasks and removing stains in both wet runs.The cube program safely halted when it could not detect a covering cup, then resumed after adding and validating a detector; for a red wedge it preserved 20 of 41 functions. The paper-towel program reused the wet-cloth wiping module, though supervised scripts were still needed to clear leftover sheets.Source: Figure 12 · p. 29, §5.5.2 · p. 30

6 Discussion

RAX frames robot learning as continual improvement through exploration, without teleoperation or parameter updates. Its accumulated knowledge is executable, verifiable, and reusable, while shared instructions and skill layers support work across robot platforms.

6 Discussion

  • RAX agents improve through accumulated experience without teleoperation or updates to their parameters.The authors contrast this approach with progress driven by larger demonstration datasets and models.Source: §6 · p. 30
  • RAX stores acquired knowledge as perception and control programs, skills, and notes that execute at robot speed, can be inspected and verified, and compose into longer behaviors.The authors expect code to play for learning from experience a role similar to network parameters in learning from data.Source: §6 · p. 30

A Interface Contracts

  • On real robots, vendor-SDK tools check commands against workspace, joint-limit, and inverse-kinematics constraints, and motion requires an armed robot.The agent's programs call the same tools; Table 11 summarizes them.Source: A Interface Contracts · p. 36, Table 11 · p. 36

B Agent Instructions and Skills

  • Every session uses a shared workspace and follows a fixed workflow: solve and verify a task, encode it as a program, test and refine it, then preserve notes and code for later sessions.Task and coordination skills are robot-agnostic, while perception and control layers provide robot-specific interfaces under shared names.Source: B Agent Instructions and Skills · p. 36, Table 12 · p. 37

C Benchmarks and Tasks

The section specifies shared simulation protocols and benchmark-specific tasks, interfaces, information, and evaluation settings. It also introduces simulation case studies tracing exploration, generated programs, and behavior on unseen scenes.

C.1 Sessions

  • Sessions begin with a coding agent and platform skills; exploration uses training-distribution scenes, while evaluation uses scenes exploration never accesses.The model weights remain unchanged, and sessions start without memory from earlier sessions.Source: §C.1 · p. 38

C.2 Rehearsal and Evaluation

  • Submitted programs are rehearsed alone in fresh environments, then evaluated without the agent or edits on unseen scenes.Only bundles that complete rehearsal without error may be submitted; both fully automated and agent-in-the-loop evaluation are specified.Source: §C.2 · pp. 38–39

C.3 RoboDojo

  • RoboDojo uses seven dual-arm tasks, with RGB-D cameras, joint-target control, and task-specific helpers; exploration layouts are separated from evaluation layouts.Agents receive task rubrics and completed-task images; the setup also describes a parallel experiment and RGB-only sessions.Source: §C.3 · p. 39, Table 13 · p. 39

C.4 BEHAVIOR-1K

  • BEHAVIOR-1K covers six household tasks with RGB-D mobile-manipulation control and evaluates on ten test instances distinct from its 300 training instances.Five tasks use fresh agent copies at evaluation, while boxing books uses a fully automated program; scoring uses the challenge Q-score.Source: §C.4 · p. 40, Table 14 · p. 40

C.5 HumanoidArena

  • HumanoidArena evaluates seven tasks on a Unitree G1 with Dex3 hands, using whole-body references and three teleoperated demonstrations visible only to the exploring agent.Exploration uses seeds excluding official seeds, and evaluation runs on 60 official test episodes per task; falls and errors count as failures.Source: §C.5 · pp. 40–41, Table 15 · p. 41

C.6 DexJoCo

  • DexJoCo tests eleven randomized-object tasks with RGB-D sensing but no force, contact, or simulator-state signals, across budgets of exploration episodes and 150 evaluation episodes.Exploration layouts use seeds 1000+i; evaluation reproduces the official layouts using seeds 0–2.Source: §C.6 · pp. 41–42, Table 16 · p. 42, Table 17 · p. 43

C.7 LIBERO and LIBERO-Pro

  • LIBERO covers 40 tasks, while LIBERO-Pro perturbs object positions; programs are evaluated on 50 official states and 50 perturbed states per task.Exploration allows up to 20 episodes on the official distribution and up to five optional adaptation episodes on the perturbed distribution.Source: §C.7 · pp. 42–43

D Simulation Case Studies

  • The simulation case studies detail what agents explore, the programs they write, and how those programs behave on unseen scenes.Source: D Simulation Case Studies · p. 44

D.1 Re-Perception Under Perturbation: Two Cans into a Basket

A program learned over 15 episodes succeeds across official and perturbed layouts by re-perceiving objects and planning placements from depth-derived geometry. Its transfer comes with a specific visibility-related failure boundary.

D.1 Re-Perception Under Perturbation: Two Cans into a Basket

  • All 50 official states and 44 of 50 perturbed states succeed after 15 exploration episodes; the perturbation moves the cans by 15 and 32 cm.The episode includes ten interactive explorations and five program rehearsals; Figure 14 depicts the layout changes and perception estimates.Source: §D.1 · p. 44, Figure 14 · p. 46

Perception

  • Gemini identifies objects, but depth supplies geometry; wrist-camera refitting checks can shape before grasping and rejects the taller, narrower ketchup bottle.The program segments depth-connected points, estimates can dimensions, and blacklists a position when the wrist refit is not can-shaped.Source: §D.1 · pp. 44–45, Figure 13 · p. 45

Placing by footprint reach

  • A wrist-camera height map with 4 mm cells supports choosing two basket spots at least 6.8 cm apart that minimize the larger can-footprint reach.The basket is remapped after the first release because its near wall hides the floor from the agent-view camera and clearance-only placements can stall on sloped walls.Source: §D.1 · pp. 44–45, Figure 13 · p. 45

Why the program transfers

  • Perturbed-layout axis estimates stay within a median 3.1 mm of ground truth and wrist refits within 1.0 mm, despite can displacements of 15 and 32 cm.All six failures share a visibility limitation: the soup is initially hidden from the agent-view camera, and the program re-perceives only from its home pose, stopping with about 300 steps left.Source: §D.1 · pp. 44–45, Figure 14 · p. 46

D.2 Self-Calibration and Grasp Design: Tower of Hanoi with Two Dexterous Hands

The dexterous-hand program completed all 150 formal Tower of Hanoi episodes on unseen layouts, while the humanoid program completed 34 of 60 formal DoubleDesk episodes. The programs combined self-measured perception with closed-loop grasping or stance planning.

D.2 Self-Calibration and Grasp Design: Tower of Hanoi with Two Dexterous Hands; D.3 From Exploration to Code: Carrying a Hammer over a Bar on a Humanoid

  • All 150 formal Tower of Hanoi episodes succeeded on unseen layouts, while the humanoid program completed 34 of 60 formal DoubleDesk episodes.The Hanoi layouts shifted the board by up to 4.9 cm; the humanoid agent predicted roughly half to two-thirds success before submission.Source: §D.2 · p. 47, §D.3 · p. 51

D.2 Self-Calibration and Grasp Design: Tower of Hanoi with Two Dexterous Hands

  • The agent measured its fingertip offsets without a hand model, then derived a rolled vertical pinch that held both disks through a 22 cm lift and 40 cm carry.The index-offset error shifted the planned pinch line by about 8 mm across a 7.7 cm disk; the adopted thumb offset was within 1.1 mm of the official model.Source: §D.2 · pp. 47–48, Figure 15 · p. 47, Figure 16 · p. 48
  • The Hanoi program measured disks and pegs on each layout, verified lifts from three cameras, retried weak pinches, and aligned disks to predicted final arm positions.Before submission, it recovered from injected alignment errors of 8 and 15 mm and a missed pinch.Source: §D.2 · p. 48
  • Across 150 Hanoi layouts, target-peg residual standard deviations were 0.24, 0.05, and 0.18 mm in x, y, and height, and the placed red disk averaged 0.66 mm below its predicted height.The maximum placement-height difference was 1.0 mm.Source: §D.2 · p. 48

D.3 From Exploration to Code: Carrying a Hammer over a Bar on a Humanoid

  • The humanoid program used depth-derived perception to locate the hammer, bar, desks, basket, and soles, and re-perceived before each stage and bar-crossing step.It verified the grasp by remeasuring the hammer handle’s height.Source: §D.3 · p. 51
  • Across 137 logged crossings, every crossing begun with the trailing sole within 35.5 cm of the bar succeeded, and the robot crossed without falling in all 59 logged formal episodes.A stance model built from 102 transitions and a three-action planner selected the crossing window.Source: §D.3 · p. 52
  • The humanoid program grew through more than thirty versions; interactive and rehearsal runs rose from 41 of 82 completions to 6 of 9 for the final versions, while grasping held in 40 of 59 logged episodes.Most formal failures occurred at the grasp.Source: §D.3 · p. 53, Table 18 · p. 53

D.4 Adaptation Under Perturbation on LIBERO-Pro

RAX resumes exploration when a program is judged unlikely to transfer to a perturbed LIBERO-Pro task. For the mirrored-cabinet task, it added a frame-reflection layer and completed all 50 perturbed states.

D.4 Adaptation Under Perturbation on LIBERO-Pro

  • RAX adapted five of 40 tasks, resuming from its workspace for at most five episodes on perturbed layouts before independent evaluation.A subagent first judged transfer from the program, task description, and first frame of a perturbed layout.Source: §D.4 · p. 55
  • The adapted cabinet program completed all 50 perturbed states after the agent added a 36-line reflection layer and narrowed its drawer-panel detector.The original program solved all 50 official states but failed all 50 mirrored states; its checks stopped it from acting blindly in 38 perturbed states.Source: §D.4 · p. 55, Figure 21 · p. 56
  • The mirror layer reflects points, orientations, and action rows across y = 0 when the cabinet is on the opposite side, with the program detecting the side from the cabinet-top height map.The layer mirrors the planning frame rather than rewriting the task program.Source: §D.4 · p. 55, Figure 21 · p. 56

E Exploration Budget of ASPIRE on LIBERO-Pro

ASPIRE develops separate programs for each LIBERO-Pro task and perturbation through execution, trace-based debugging, and validation. Its Long-suite result uses a different setting: a LIBERO-90 skill library without debugging on Long tasks.

E Exploration Budget of ASPIRE on LIBERO-Pro

  • ASPIRE writes programs in CaP-X with Claude Code and Claude Opus 4.6, using task-and-perturbation-specific development and interfaces for perception, grasp planning, inverse kinematics, and joint-space motion.Its programs may not read simulator ground-truth state.Source: E Exploration Budget of ASPIRE on… · p. 57
  • ASPIRE executes an initial program on 15 debugging seeds, uses trace-informed repairs with up to three replays per seed, validates on 15 further seeds, then evaluates on seeds 1–50.Evolutionary search is used on some tasks and can run up to four iterations; the number of candidates K is not reported.Source: E Exploration Budget of ASPIRE on… · p. 57
  • With at most five evaluation-distribution rollouts on five adapted tasks and none on the other 35, RAX exceeded ASPIRE under position perturbation by 28.2 points on Spatial, 2.0 on Object, 6.0 on Goal, and 48.8 on Long.ASPIRE’s Long result came from applying a library built on LIBERO-90 without debugging on Long tasks.Source: E Exploration Budget of ASPIRE on… · p. 57
  • ASPIRE’s debugging and validation rollouts come from the same perturbed-suite distribution as evaluation, and its total rollouts per task are not reported.The task perturbation can change the goal, so a program developed on the original task would pursue the original goal.Source: E Exploration Budget of ASPIRE on… · p. 57, Table 20 · p. 57

F Summary Results of Figure 1

Figure 1 compares RAX with the strongest prior learned policy models across five simulation benchmarks. RAX’s mean margin is 37.4 points, while the per-budget chart averages benchmark success rates with equal benchmark weight.

F Summary Results of Figure 1

  • RAX’s average margin over the strongest prior models was 37.4 points across RoboDojo, BEHAVIOR-1K, HumanoidArena, bimanual DexJoCo, and LIBERO-Pro.The benchmark margins were 53.4, 38.3, 7.9, 45.5, and 42.1 points, respectively.Source: F Summary Results of Figure 1 · p. 58
  • With 32 exploration episodes, RAX averaged 94.3% on the five bimanual DexJoCo tasks, versus 48.8% for ME-Dex.The chart excludes six single-arm DexJoCo tasks because RAX solves them within a few exploration episodes.Source: F Summary Results of Figure 1 · p. 58
  • The per-budget chart averages task success rates within each benchmark, then gives the five benchmarks equal weight; a task missing a shown budget uses its largest smaller-budget result.The dashed 41.2% line is the mean of the strongest prior models in the right-hand chart.Source: F Summary Results of Figure 1 · p. 58
  • The 37.4-point comparison is against learned policies, whereas comparisons with the strongest prior result of any kind use different margins, including +31.4 on all eleven DexJoCo tasks and +21.2 on LIBERO-Pro.The all-task DexJoCo comparison is 96.7% for RAX versus 65.3% for ME-Dex at 32 exploration episodes.Source: F Summary Results of Figure 1 · p. 58

G Real-Robot Tasks

The real-robot experiments used six tasks across two robot setups, with agents given task text and robot tools but no demonstrations, object models, or trained policies.

G Real-Robot Tasks

  • Five tasks used a Franka FR3 with a parallel gripper and three RGB-D cameras; the sixth used a Rainbow RB-Y1 mobile manipulator with a head RGB-D camera.The agent could request operator help through tools and chat; resets and interventions were recorded.Source: G Real-Robot Tasks · p. 59

G.1 Find the Cube

Across these real-robot tasks, agents explored and revised programs to handle changing task conditions. The air-fryer program used scene-specific 3D estimates and collision checks, while the cube task adapted to new hiding cases and a different object.

G.1 Find the Cube

  • After the cube-hiding method failed to recognize an upright-cup case, the agent added a detector and resumed successfully; it later grasped a red wedge on the first attempt.The resumed cube trial took 98 s; for the wedge, 20 of 41 program functions remained unchanged.Source: §G.1 · p. 59, Figure 22 · p. 60, Figure 23 · p. 61

G.2 Air Fryer

  • Four clean unattended air-fryer successes included a moved and a rotated layout; the final run placed two objects and closed the drawer in 210 s.The handle finished 3.5 mm from its initial position.Source: §G.2 · p. 63, Figure 25 · p. 64, Figure 27 · p. 67

G.2.1 Agent-Written 3D Perception

  • For the air fryer, the agent fitted robust estimates of its axis, handle direction, handle geometry, banana, and basket rim from depth points.It checked each planned motion against an articulated fryer model and verified the drawer opening, object placement, and final handle position with new measurements.Source: §G.2.1 · pp. 65–66
  • The agent adapted its air-fryer perception to sensor limits and a camera bump by using exterior views for metric targets and re-estimating the bumped camera pose.It stopped using the other camera after it moved too far to correct.Source: §G.2.1 · pp. 63, 68

G.3 Wiping a Stain

The agent developed a towel-wiping program and adapted its reusable components to paper towels, while some cleanup in the paper-towel trials still required supervised scripts.

G.3 Wiping a Stain

  • The towel was pressed 12 mm below measured contact height without force sensing, and the stain varied in size, shape, and position.The agent measured stain areas from 10 to 311 cm2.Source: §G.3 · p. 69
  • Trials 6 and 7b cleaned the stain in one pass without intervention, with 0 cm2 residue measured in both exterior and wrist cameras.The program located the stain, wiped back and forth, and checked for residue.Source: §G.3 · pp. 69, 71, Figure 28 · p. 69, Figure 30 · p. 70
  • The paper-towel variant reused the towel program’s stain localization, wiping-region fit, wad handling, zig-zag wipe, and residue check.After three dry runs, both wet runs removed the stain, but leftover sheets and a flat wet scrap needed supervised cleanup scripts.Source: Figure 29 · p. 69, §G.3 · p. 71

G.4 Peg Insertion: the ManipulationNet Peg-in-Hole Challenge

The agent explored peg insertion with user supervision and completed a full run placing three pegs in the tight L3 holes using tilted entry and force-guided adjustments.

G.4 Peg Insertion: the ManipulationNet Peg-in-Hole Challenge

  • The task used three unchamfered pegs and holes with clearances down to 0.02 mm, but the agent was asked to progress only through L3.L4 was not attempted; the board was taped down after it moved during a jammed-peg lift.Source: §G.4 · p. 72
  • All three pegs seated in L3 during the full run, at depths of 20.8 mm, 19.7 mm, and 21.0 mm out of 21.8 mm.The run took about 55 minutes, including operator extraction of pegs from earlier attempts.Source: §G.4 · p. 72, Figure 31 · p. 72
  • For L3, the agent tilted a peg corner-first, then turned it upright while making small rocking pushes.The procedure registered the board from CAD and wrist-camera image, then used touch to find holes.Source: §G.4 · p. 72, Figure 32 · p. 73

G.5 Bandu Tower

The Bandu and RB-Y1 trials show RAX completing precision assembly and a six-subtask mobile sequence through iterative program development and skill reuse. The trials also document operator involvement and checks used to assess completion.

G.5 Bandu Tower

  • The Bandu task required rebuilding an irregular seven-piece tower, including placements with little clearance and an estimated 1 mm opening for the knob.The initial tower served as the target; the agent was not shown how it had been built.Source: §G.5 · p. 74
  • The Bandu operator reset pieces once, re-homed the gripper twice, moved an unreachable piece, and requested visual knob alignment rather than contact search.The requested alignment allowed at most 1 mm of opening error.Source: §G.5 · p. 74
  • The final Bandu program disassembled and rebuilt all 7 pieces automatically; the knob was aligned to a 0.19 mm offset before insertion.Each placement was checked by seat height, with an additional uprightness check for the urn.Source: §G.5 · p. 74

G.6 Long-Horizon Tidying on the RB-Y1

  • The RB-Y1 session reused skills from five earlier sessions and added a reaching skill and a whole-task composition skill.No existing skill covered carrying the cups and banana into the basket.Source: §G.6 · p. 76
  • During RB-Y1 trials, the operator reset the scene and advised the agent; after a cabinet collision, the agent checked the base’s swept footprint before each base move.The operator also re-armed the robot after controller faults in trial 3.Source: §G.6 · p. 76
  • All six RB-Y1 subtasks were completed in each of three trials; trial 3 took 98 minutes with the agent at every skill checkpoint.The agent judged each subtask’s success from head-camera images.Source: §G.6 · p. 76

H Compute and Cost

Model calls across the paper’s simulated and real experiments consumed over 100k US dollars in API credits.

H Compute and Cost

  • Across all simulated and real experiments, model calls consumed over 100k US dollars of API credits.This is the reported aggregate cost of model calls across the paper’s experiments.Source: H Compute and Cost · p. 77

Related papers

Loading 2610.22496v1…