NVIDIA Open-Sources Robot Agent Framework ASPIRE 1.0

NVIDIA has officially open-sourced ASPIRE V1.0, a robot agent framework that enables robots to autonomously write control programs, execute tasks, review failures, modify code, and store validated experience into a skill library. The code, experiment configurations, and skills are all publicly available.

ASPIRE, which stands for Agentic Skill Programming through Iterative Robot Exploration, treats the robot as an agent capable of running and debugging code. During task execution, the system records multimodal trajectories including vision results, candidate grasp points, motion paths, and collision feedback. When a task fails, the agent traces the trajectory to locate the issue, modifies the control program, and re-executes while exploring alternative strategies. Effective approaches are then consolidated into skills within an expanding library.

The current skill library contains over 90 skills across 150+ tasks in 18 categories, covering grasping, navigation, object search, perception, bimanual manipulation, and failure recovery. These skills can be reused across tasks and transferred from simulation to real robots, improving first-try debugging efficiency even when robot morphology and APIs change.

According to NVIDIA's robotics lead Jim Fan, after switching the underlying model to OpenAI's GPT-6 Astra on its release day, ASPIRE achieved a 99% success rate on 80 tasks in LIBERO-Pro and 100% on representative tasks in Robosuite and BEHAVIOR-1K. Fan described this as a "Physical Bitter Lesson," noting that general-purpose models' reasoning and coding capabilities are entering the physical world through such frameworks.