RAX: Scaling Online Exploration Beats SOTA by 37.4% Across 5 Robot Benchmarks

Introduction: The Challenge of Scaling Robot Learning

Robotics has pursued scaling through more teleoperation data and larger models (VLA, WAM), yet deployment on new robots or environments still demands massive human data collection and retraining. Thousands of pre-training hours have not made zero-shot robust deployment routine. The authors argue that true embodied intelligence should come from continuous online exploration and self-improvement at test time, similar to human skill acquisition.

RAX: Scaling Online Exploration for Robot Agents

A joint team from Harvard, NVIDIA, UC Berkeley, MIT, and Stanford proposes RAX (Robot Agent eXploration), which brings Scaling Online Exploration to robot learning. A general-purpose coding agent connects directly to the robot's perception and control interfaces, explores to gain experience, writes successful decisions as executable programs, and accumulates them as reusable skills. Model parameters remain frozen throughout.

From Scaling Data to Scaling Exploration

RAX mimics human practice: try, observe, correct, and retain effective methods. The agent receives RGB images, depth, camera parameters, and force feedback, then autonomously writes perception and estimation programs — converting depth to 3D point clouds, extracting object geometry, estimating poses and affordances — and feeds these into action planning. The robot simultaneously learns how to perceive and how to act for the task.

Air Fryer Example: Autonomous Problem Solving

In an air fryer task, the agent builds geometric models of the basket and drawer from depth, estimates handle orientation, and searches for grasp and pull strategies. For a 18 cm banana into a 19 cm basket, it discovers a 60-degree tilt release. When pulling a long handle causes IK limits, it switches to gripping the shorter handle junction. These solutions emerge from exploration and planning, exceeding typical human teleoperation capabilities.

Code as Policy: From Exploration to Reusable Skills

After successful exploration, the agent organizes perception, planning, and control into standalone programs ( Code as Policy ). In new environments, the agent continues exploring, directly modifying programs to adapt — no policy training or fine-tuning needed. Programs re-localize objects, generate actions, and verify key steps (grasp confirmed, drawer opened, object placed, retry needed). A successful method from one exploration becomes a recomputable, re-executable capability for future scenes.

When programs encounter uncertain perception or planning issues, they raise exceptions back to the agent, which inspects, diagnoses, modifies the module, and resumes testing. As exploration progresses, more decisions are offloaded to fast program execution while the agent handles novel variations.

Cross-Embodiment Generalization via Unified Interface

RAX uses a front-end/back-end architecture. The front end handles exploration, program generation, memory, and skill organization; the back end provides perception, control, and low-level execution. The same agent framework connects single-arm, dual-arm, dexterous hands, mobile robots, and humanoids. New embodiments are integrated via interfaces; the agent learns tasks through interaction without prior demonstration collection or policy retraining. Across different joint structures, workspaces, and sensor configurations, the agent discovers embodiment-specific perception and control methods, saving successes as programs. This shared exploration ability lets robots learn tasks according to their own bodies, enabling a continuously scalable platform for cross-embodiment robot learning.

Benchmark Results: 37.4% Average Lead Over SOTA

On five simulation benchmarks — RoboDojo (hard subset), BEHAVIOR-1K (hard subset), HumanoidArena, DexJoCo, and LIBERO-Pro — RAX leads the previous best SOTA by an average of 37.4 percentage points . Specific gains:

RoboDojo hard: from 24.9% to 78.3%

BEHAVIOR-1K hard: from 8.2% to 46.5%

DexJoCo dual-arm: from 48.8% to 94.3%

Crucially, RAX uses zero teleoperation demonstrations , while compared VLA/WAM models use 50–100 demonstrations per task. On original LIBERO suites, RAX achieves 98.5% average success. After object position shuffling, LIBERO-Pro adaptation within 5 episodes reaches 84.4%, far above VPP2 (42.3%) and ASPIRE (63.2%). On difficult tasks, prior policy models scored 0 on RoboDojo pouring and block swap; RAX achieves 100%. On BEHAVIOR-1K long-horizon packing (10 minutes), all previous methods scored 0 (partial 0.2); RAX achieves 100%. Exploration scaling shows clear capability jumps: charger insertion rises from 26.0% (4 episodes) to 94.0% (8 episodes); block swap from 34.0% (4 episodes) to 100% (16 episodes).

Real-World Robot Experiments: Zero Demonstrations, Frozen Model

On physical robots (including RB-Y1), RAX masters precision peg-in-hole (0.1 mm clearance), air fryer operation, stain wiping, hidden object tracking, and irregular block tower building — all without any human demonstrations. For the transparent peg-in-hole, the agent actively contacts surfaces with the arm, using force feedback to measure height and correct localization. The robot decides how to acquire needed information.

Long-horizon mobile whole-body control demonstrates skill composition: RB-Y1 moves a chair, stacks cups, places cup and banana in basket, transports basket to cabinet, returns chair, and carries a box — six stages invoking previously practiced skills, stitched with live perception. Throughout, zero teleoperation demonstrations and zero policy training/fine-tuning; model parameters stay frozen.

When environments or tasks change, the agent retains valid modules and updates only necessary parts. Wiping programs transfer to tissue manipulation; transport, grasp, and chair-moving skills combine into longer chains. Each exploration leaves programs that become starting points for future tasks, enabling continuous experience accumulation and capability expansion.

Programmatic skills drastically cut inference latency: air fryer task inference drops from ~72 minutes during initial exploration to 18 seconds for standalone program execution — a 241× reduction . Exploration-time thinking is saved as fast, callable execution capability.

Open Source Release

The team releases the complete RAX framework, agent instructions, general workflows, robot interfaces, benchmark integration, evaluation protocols, and all real-robot exploration session logs (control programs and skills). Researchers can follow the same pipeline: connect a robot, launch exploration, inspect execution, and save successes as reusable capabilities. The platform supports adding new embodiments and tasks, studying how more exploration converts into continuously growing robot abilities, and provides a unified interface for advancing general agents into the physical world.