NVlabs/ASPIRE

What it solves

ASPIRE addresses the challenge of training robots by shifting the paradigm from traditional gradient-based learning to a process of "skill discovery" and refinement. Instead of updating model weights through standard training, it treats a robot's intelligence as a repository of sensorimotor skills that can be iteratively improved by agents.

How it works

ASPIRE uses coding agents (such as Claude Code or Codex) to manage the robot's learning process. These agents utilize specific "skills"—such as iterative debugging and evolutionary search—to refine robot behaviors in simulation environments like LIBERO-Pro, Robosuite, and BEHAVIOR-1K. For real-world tasks, it employs a "System 2" operator runbook that guides the agent through complex manipulation tasks like solving a Rubik's Cube or unscrewing a bottle, relying on language-model-generated code to handle grasping, recovery, and verification.

Who it’s for

This project is designed for robotics researchers and developers interested in agentic workflows for robot learning. It is particularly useful for those looking to scale robot performance across diverse tasks without relying on massive, fixed datasets or traditional fine-tuning.

Highlights

• Achieves high success rates (up to 99-100%) on complex benchmarks like LIBERO-Pro and BEHAVIOR-1K.
• Enables cross-embodiment sim-to-real transfer for long-horizon manipulation tasks.
• Implements an agent-agnostic framework where the "training" process is handled by coding agents executing iterative debugging or evolutionary search.
• Provides a structured approach to real-world robot manipulation, including recipes for tasks like cup stacking and pill insertion.