NVIDIA Introduces ASPIRE — Self-Learning Robotics Framework Achieving 31% Zero-Shot Success on Complex Tasks

NVIDIA AI on July 3, 2026 presented ASPIRE — a framework for robotic control that autonomously generates robot control programs, iteratively corrects errors, and stores verified solutions in a reusable skill library. According to NVIDIA, on the LIBERO-Pro benchmark, the system gained up to 77 points and achieved 31% success on new long-horizon tasks in zero-shot mode — without additional training on these scenarios.

How the self-improvement cycle works in ASPIRE

The core of ASPIRE is an iterative cycle of generation and self-correction: the framework writes a robot control program as code, runs it, detects failures, and makes corrections until the task is completed successfully. Code generation as a tool for robotic control is already used in research, but ASPIRE takes a principled step further: successful repairs are "distilled" into a structured skill library for reuse.

The library works as the system's long-term memory. When ASPIRE encounters a task similar to one already solved, it accesses ready-made debugged blocks — instead of starting from scratch. This reduces the number of attempts to successful completion and allows the system to transfer accumulated experience to unfamiliar scenarios. The accumulated library grows: each new task potentially adds new patterns to it or refines existing ones.

• ASPIRE publication date — July 3, 2026
• Gain on LIBERO-Pro benchmark — up to 77 points
• Zero-shot accuracy on LIBERO-Pro Long Tasks — 31%
• Transfer of skills to tasks not included in the training set

Why LIBERO-Pro Long Tasks is a complex target?

LIBERO-Pro is a recognized benchmark for evaluating robotic systems on tasks with long planning horizons. Unlike simple single-step tasks, long-horizon tasks require sequential execution of a multi-stage chain: find the needed object, move it to the target, open a container, place the object and close it. An error in any link — failure of the entire episode.

LIBERO-Pro Long Tasks is the most complex part of the benchmark with the longest sequences. Zero-shot metric means the system performs the task for the first time: without demonstrations, without additional training on the specific scenario. On such tasks, basic methods often give results close to zero. 31% success rate in zero-shot mode is a non-trivial indicator for this class of systems.

The gain up to 77 points on LIBERO-Pro demonstrates the gap between ASPIRE and baseline methods on standard tasks from the same benchmark.

Where ASPIRE fits in robotic AI

The ASPIRE approach reflects a broader trend: the use of language models for controlling physical systems through code generation. Unlike classical reinforcement learning, which requires millions of simulations, ASPIRE relies on iterative code generation — an approach economical in terms of environment interactions.

The key innovation — an explicit skill library — solves a long-standing problem of robotic systems: accumulating experience without losing it when transitioning to new tasks. Unlike neural network approaches, where knowledge is stored implicitly in model weights, ASPIRE's library is structured and accessible for expansion.

What this means

ASPIRE offers a model in which a robotic agent gradually accumulates experience in the form of reusable software blocks — instead of solving tasks from scratch each time. The NVIDIA AI publication fits into the search for a path from robots working strictly within the training distribution to systems capable of generalizing experience to new scenarios. If the approach proves scalable in real conditions, industrial robots will be able to improve their capabilities directly during operation — without constant manual retraining.