NVIDIA unveils ASPIRE, a self-improving robotics framework that hits 31% zero-shot on long-horizon tasks

On July 3, 2026, NVIDIA AI unveiled ASPIRE, a framework for robotic control. ASPIRE generates robot control programs on its own, fixes errors iteratively, and saves verified solutions to a reusable skill library. According to NVIDIA, the system gained up to 77 points on the LIBERO-Pro benchmark and reached a 31% success rate on new long-horizon tasks in zero-shot mode, without any additional training on those scenarios.

How ASPIRE's self-improvement loop works

At the core of ASPIRE is an iterative loop of generation and self-correction. The framework writes a robot control program as code, runs it, logs failures, and applies fixes until the task succeeds. Researchers already use code generation to control robots, but ASPIRE goes a significant step further: it distills successful repairs into a structured skill library for reuse.

The library serves as the system's long-term memory. When ASPIRE encounters a task similar to one it has already solved, it draws on ready-made, debugged building blocks instead of starting from scratch. That cuts the number of attempts needed to succeed and lets the system carry accumulated experience over to unfamiliar scenarios. The library keeps growing: each new task can add new templates or refine existing ones.

• ASPIRE publication date: July 3, 2026
• Gain on the LIBERO-Pro benchmark: up to 77 points
• Zero-shot accuracy on LIBERO-Pro Long Tasks: 31%
• Skill transfer to tasks outside the training set

Why is LIBERO-Pro Long Tasks a hard target?

LIBERO-Pro is an established benchmark for evaluating robotic systems on long-horizon planning tasks. Unlike simple single-step tasks, long-horizon tasks require the robot to execute a multi-stage chain in sequence: find the right object, move it to the target, open a container, place the object inside, and close it. A mistake at any link fails the entire episode.

LIBERO-Pro Long Tasks is the hardest part of the benchmark, with the longest sequences. The zero-shot metric means the system attempts the task for the first time: no demonstrations and no fine-tuning on the specific scenario. Baseline methods often score close to zero on these tasks. A 31% zero-shot success rate is a nontrivial result for this class of systems.

The gain of up to 77 points on LIBERO-Pro shows the gap between ASPIRE and the original baseline methods on the benchmark's standard tasks.

Where ASPIRE fits in robotic AI

ASPIRE's approach reflects a broader trend: using language models to control physical systems through code generation. Classic reinforcement learning requires millions of simulations; ASPIRE instead relies on iterative code generation, an approach that needs relatively few interactions with the environment.

The key innovation, an explicit skill library, tackles a long-standing problem in robotics: accumulating experience without losing it when moving to new tasks. In neural network approaches, knowledge lives implicitly in the model weights. ASPIRE's library, by contrast, is structured and open to extension.

What this means

ASPIRE proposes a model in which a robotic agent gradually builds up experience as reusable program blocks instead of solving every task from scratch. The NVIDIA AI release fits into the broader search for a path from robots that work strictly within their training distribution to systems that can generalize experience to new scenarios. If the approach proves scalable in real-world conditions, industrial robots could improve their capabilities directly during operation, without constant manual retraining.