Robotics

ASPIRE: Agentic /Skills Discovery for Robotics

Runyu Lu, Yubo Wu, Ethan Kou, Letian Fu, Wenli Xiao, Ajay Mandlekar, Yinzhen Xu, Guanya Shi, Ken Goldberg, Ang Chen, Mosharaf Chowdhury, Yuke Zhu, Linxi Fan, Guanzhi Wang

Featured July 19, 2026

Original Paper

AI-generated analysis — This is SciGrove's AI interpretation of the paper, not peer-reviewed content. Always refer to the original paper.

Simply

Robots usually struggle with complex tasks and learning from mistakes, but this system helps them write and fix their own programs by carefully watching what went wrong, saving good solutions in a skill library, and trying many different ways to solve problems.

In depth

The paper introduces Aspire, a continual learning system for robotics that autonomously writes and refines robot control programs. It achieves this by using a closed-loop robot execution engine that provides fine-grained multimodal traces for failure diagnosis, a continually expanding skill library to store validated repairs, and an evolutionary search procedure to generate diverse task sequences and programs, moving beyond single-trajectory refinement. This allows robots to learn and adapt more effectively over time, improving performance and enabling zero-shot generalization and sim-to-real transfer.

Key Takeaways

1
Aspire autonomously writes and refines robot control programs using a code-as-policy paradigm, significantly improving performance on complex manipulation tasks.
2
The system features a continually expanding skill library that distills validated failure fixes into reusable, transferable robotic knowledge, enabling zero-shot generalization.
3
A closed-loop robot execution engine provides fine-grained multimodal traces for precise failure diagnosis, complemented by an evolutionary search for diverse program exploration.

Conceptual Flow

HIGH LEVEL

Methodology: How Aspire Learns

The system helps robots learn by letting them write code, try it, see detailed mistakes, save good fixes, and then try many new ideas to get better.

Robot Task Goal

Write Code

Try Code

See Mistakes

Save Good Fixes

Try New Ideas

Results: Impact of Aspire

By learning this way, the robot gets much better at tricky tasks, can do new tasks it hasn't seen before, and even use what it learned in computer worlds to help real robots.

Old Robot Learning

New Smart Learning

Better Task Success

New Task Skills

Sim-to-Real Help

This breakdown was generated by SciGrove. Get the same analysis — intuition, storyboard, peer review, a runnable prototype and a glossary — on any paper you upload or paste a DOI for.