Most robot control pipelines rely on hand-crafted modules or end-to-end learned policies that are brittle under distribution shift. ASPIRE takes a different perspective: treat program text as the policy, use coding agents to propose and iteratively debug control programs, and distill validated fixes into a growing library of reusable skills. This lets the system compund experience across tasks, embodiments, and sim/real settings rather than retraining monolithic weights for each new scenario.
What Sets It Apart
- Code-as-policy continual learning: agents write executable Python control programs, run them in a closed-loop execution engine that exposes multimodal traces (perception overlays, trajectories, collisions), diagnose failures, and synthesize repairs. So what: faults are fixed at the program level and distilled into modular skills rather than buried in opaque weights.
- Skill library that compounds experience: discovered fixes become reusable primitives that accelerate zero-shot generalization and transfer between tasks, simulators, and different robot embodiments. So what: library growth yields faster adaptation and stronger zero-shot performance on unseen long-horizon tasks.
- Evolutionary, multi-agent search plus closed-loop verification: an evolutionary search generates diverse candidate programs while iterative debugging validates outcomes and extracts robust repairs. So what: exploration goes beyond single-trajectory refinement, improving robustness under perturbations and long-horizon failure modes.
- Demonstrated scale and sim-to-real: achieves near-saturation on standard benchmarks with strong LLMs (e.g., 99% success on 80 LIBERO-Pro tasks with Claude Opus 5.5), provides sim-to-real skill transfer and online real-robot learning (including a Rubik’s Cube demo). So what: empirical evidence that coding-agent-driven skill discovery can reach high success rates on complex manipulation benchmarks and reduce real-robot programming effort.
Who It's For and Trade-offs
Great fit if you need a research-grade platform to explore agentic program synthesis for robotics, want modular reusable manipulation skills, or are reproducing long-horizon household and benchmark experiments (LIBERO-Pro, Robosuite, BEHAVIOR-1K). It is also suited for teams that can run closed-loop simulation and accept generated code execution under strict isolation.
Look elsewhere if you require turnkey industrial deployment without manual oversight: the system runs language-model-generated Python (security risks), real-robot runs require operator controls and careful sandboxing, and reproducing top results depends on access to high-capability coding agents and compute. ASPIRE focuses on autonomous skill discovery and research reproducibility rather than packaged production integration.
