ASPIRE is a self-improving continual learning system for robotics that writes and refines code-as-policy robot control programs from execution feedback. It inspects rollout traces, diagnoses failures, repairs programs, validates corrected behaviors, and saves reusable skills for future tasks. ASPIRE operates in an open-ended learning loop with three key components: a closed-loop robot execution engine, a continually expanding skill library, and an evolutionary search procedure.
For each perception, planning, grasping, and control call, the engine records the observations, inputs, outputs, and visual evidence if possible. These rich multimodal traces allow the agent to selectively inspect salient primitive logs, progressively localize failures, and validate repairs through re-execution.
ASPIRE maintains a growing skill library that distills validated fixes into modular, transferable robotic knowledge retrievable as in-context guidance for future tasks.
ASPIRE employs an evolutionary search procedure that generates diverse task sequences and control programs, exploring beyond single-trajectory self-improvement through iterative debugging and parallel refinement.
Each card shows either a real-robot example of ASPIRE or a CaP-Agent0 baseline or initial task attempt paired with the final behavior after ASPIRE diagnosed the rollout and rewrote the program. Hover over a demo video to view the baseline or initial program and the ASPIRE fix code. Use the filters to narrow the gallery. This is not an exhaustive list of all tasks.
Explore the expansive skill library learned by ASPIRE. Each card groups many related skills. This is a representative sample, not an exhaustive list.
We tested ASPIRE across a full suite of recent frontier models, including GPT-6 Astra, Claude Opus 5.5, Claude Sonnet 5.5, Claude Fable 5.1, and open-weight models such as Kimi K3 and Qwen 3.8 Max, while holding the environment, robot programming API, and evaluation protocol fixed.
Evaluation is run on 10 LIBERO-Pro tasks.
ASPIRE also runs online on a real robot: the coding agent acts as a system-2 that calls executable perception and action tools and records what it learns as text skills.
Tokens and wall-clock time per completed face turn fall across episodes as context and skills accumulate. Values are full-episode amortized, including failed attempts, recovery, and context compaction.
Each video compares the human teleoperation demonstration from the original LIBERO dataset with a held-out evaluation episode of the ASPIRE program written by each model: Claude Opus 4.6, Claude Opus 5, Claude Opus 5.5, and GPT-6 Astra.
We show ASPIRE with Claude Opus 4.6, GPT-6 Astra, and Claude Opus 5.5 (Opus 5.5 was not evaluated on BEHAVIOR-1K). ASPIRE learns on a small debug split and reports success on disjoint held-out evaluation seeds, while CaP-Agent0 regenerates a program for each seed with test-time reasoning and retries.
LIBERO-Pro Object / Goal / Spatial under position perturbations (shifted object positions) and task perturbations (perturbed instructions): 160 tasks, each evaluated on 50 held-out seeds.
ASPIRE handles contact-heavy tabletop manipulation tasks. Each task is evaluated on 100 held-out seeds.
BEHAVIOR-1K tests long-horizon mobile manipulation. Each task is evaluated on 25 held-out seeds.
Skills learned across LIBERO-90 carry over to held-out long-horizon tasks. As the skill library grows, coding agents achieve higher success rate on LIBERO-Long.
For each task, we compare real-robot debugging with and without retrieving a corresponding simulation-discovered skill from ASPIRE’s library. Token counts are measured until the first successful real-robot program. Success Rate reports held-out evaluation trials of the code that achieved that first success. When equipped with ASPIRE skills, the coding agent reaches its first success with fewer tokens, and the generated code achieves a higher success rate.
Output tokens (M)
Total tokens (M)
Held-out success rate (%)
LIBERO-Pro Goal / Spatial / Object / Long under task and position perturbations: 160 tasks, each evaluated on 50 held-out seeds.
We present ASPIRE, a self-improving continual learning robotic system that autonomously writes and refines robot control programs while compounding experience into a reusable skill library. ASPIRE operates in an open-ended learning loop with three components: a closed-loop robot execution engine that exposes fine-grained multimodal traces, a continually expanding skill library that distills validated fixes into transferable knowledge, and an evolutionary search procedure that explores diverse task sequences and control programs. Across diverse benchmarks, ASPIRE substantially outperforms existing VLA and coding-agent baselines, demonstrates strong zero-shot transfer to unseen long-horizon tasks, and provides initial evidence that the skills discovered in sim can transfer across embodiment to significantly reduce real-robot programming token cost despite different robot embodiments and APIs.
*Equal contribution · †Project leads
@article{lu2026aspire,
title = {ASPIRE: Agentic /Skills Discovery for Robotics},
author = {Runyu Lu and Yubo Wu and Ethan Kou and Haotian Lin and Letian Fu and
Wenli Xiao and Ajay Mandlekar and Yinzhen Xu and Guanya Shi and
Ken Goldberg and Ang Chen and Mosharaf Chowdhury and Jan Kautz and
Yuke Zhu and Linxi "Jim" Fan and Guanzhi Wang},
year = {2026},
journal = {arXiv preprint arXiv:2607.00272},
url = {https://arxiv.org/abs/2607.00272}
}
We thank Nadun Ranawaka, Jimmy Wu, Matin Furutan, Abhi Maddukuri, Yulu Gan, Matin Nikoui, and Yuqi Xie for their help with open-source support, real robot infrastructure, advice and guidance on the paper, website release, and experimental equipment. We are also grateful to the members of NVIDIA GEAR, UMich SymbioticLab, UsesysLab, UC Berkeley AUTOLab, CMU LeCAR Lab for their kind support.