PyRUA-REAL: Robot Exploration and Adaptive Learning through RSI in the Real World

PyRUA-REAL uses researcher–engineer coding agents with persistent experience to develop and freeze real-world robot programs, reporting 91.1% average success on eight tasks and 84.42% on 42 RoboDojo tasks.

Coding agents can translate instructions into robot programs, but it is unclear how physical interaction experience can become persistent robot capability that improves across attempts. Real-world exploration requires deciding which experiments to run, revising implementation details, and recognizing when to change the underlying approach. Execution feedback can be ambiguous, and success under one configuration does not establish robustness or reuse without online agent decisions.

PyRUA-REAL is an embodied RSI framework with improvement, execution, and experience components. A researcher agent selects objectives, technical approaches, stopping criteria, and relevant experience; an engineer agent implements programs, runs robot trials, diagnoses failures, and returns structured reports. Experience is stored at three levels: facts, memory, and implementation. Developed task programs execute independently with live perception, planning, control, and recovery routines, and can invoke locally deployed pretrained perception models. The study evaluates eight real-world tasks on a dual-arm Franka platform and 42 RoboDojo tasks.

• Across eight real-world tasks, PyRUA-REAL achieves an average success rate of 91.1%; on 42 RoboDojo tasks it achieves an average success rate of 84.4%. §Abstract
• Agent-free real-world programs were produced for seven of the eight tasks, with recorded development success from 91.67% to 100%; time to first success ranged from 1.5 hours to 45 hours, and Ethernet insertion remained unsuccessful after ten days without online assistance, then succeeded after approximately 15 additional hours. §4.3.1
• For five-block stacking, median stack–unstack cycle time fell from 285.3 to 156.0 s, mean lateral error from 2.69 to 0.44 mm, mean absolute angular error from 1.37 to 0.25 degrees, and full-cycle success rose from 88.7% to 98.7% across two reported smooth-stage evaluations. §4.3.2
• Tic-tac-toe success rate rose from 50% in the first version to 100% in the next, and game-and-reset time fell from 577 to 535 s; between V4 and V5 time dropped from 573 to 536 s when transfers switched to a single JointMotion. §4.3.2
• Hanoi median stack–unstack cycle time reached 1402 s in the robust stage, then fell to 883 and 793 s in speed-focused versions; full-cycle success rose from 60% initially to 100% in the final version. §4.3.2
• On 42 RoboDojo tasks, frozen programs achieved 84.42% capability-averaged success versus 50.35% for PyRUA-Lean and 22.48% for GPT-6-Astra; Precision was 74.00% versus 38.25%, Long-Horizon 56.00% versus 43.75%, and Generalization 41.17% versus 43.17%. §4.3.3

• Ethernet insertion remained unsuccessful without online agent participation in the reported development record (§4.3.1).
• Real-world success rates aggregate attempts across program revisions and describe the development record, not single frozen-program reliability (§4.3.1).
• Placement errors are visual estimates, not independent measurements of physical accuracy (§4.3.2).
• RoboDojo comparisons are reported results from separate evaluations with different or unspecified seeds, not a matched-seed comparison; the Introduction reports 84.82% while §4.3.3 and Appendix D report 84.42%.

• A researcher agent selects objectives, approaches, and stopping criteria, while an engineer agent implements, tests, and diagnoses within structured experiments; reports return results, learned findings, evidence, and code. §3.3
• Experience is organized into L0 facts, L1 memory, and L2 implementation; the researcher retrieves facts and memory, while the engineer inspects, adapts, and tests code. §3.4
• Seven of eight real-world tasks produce agent-free programs; Ethernet insertion remains unsuccessful without online agent participation in the development record. §4.3.1
• Continued improvement after first success is traced for stacking, tic-tac-toe, and Hanoi, with different trajectories, occasional regressions, and shifts between reliability and speed. §4.3.2
• RoboDojo frozen programs reach 84.42% average success, with gains concentrated in Precision and Long-Horizon while Generalization is slightly lower than PyRUA-Lean. §4.3.3
• Explicit task requirements are evaluated on hidden layouts in Table 4; per-task before/after success varies, with some tasks unchanged and Swap blocks negative. §B
• Primitive-skill transfer benefits five of the eight target tasks when agents have access to skills acquired on earlier tasks. §4.7
• Offline coding freezes programs for deployment without language-model calls, while deployment-time coding includes agent calls; the comparison excludes development time. §4.8