Log in Sign up
Opinion He Sold Everything and Rode a Bus 2,000 km to Marry an AI Farm Girl With 300,000 Followers Business Anthropic Pledges 5 Percent of Itself to a Public Interest Fund That Grows With Its Valuation AI Agents NVIDIA Open-Sources ASPIRE, the Robot Framework That Turns Failed Grasps Into a Growing Skill Library Coding Assistants PostgreSQL 19 Slips to November 19 After LLM-Assisted Code Review Turned Up Forty Defect Reports News VAST Open-Sources EdenLive, a 5B Real-Time World Model That Tops 24 FPS at 720p on a Single RTX 5090 Business Nvidia Tells Samsung, SK Hynix and Micron to Build Its Custom HBM Base Dies on TSMC's 3nm Process Claude At Anthropic, Junior Share of Technical Staff Fell From 21 Percent to 14 Percent in a Year News Holyvolt Raises 45 Million Dollars at a 245 Million Valuation to Let AI Pick the Next Battery Chemistry Opinion He Sold Everything and Rode a Bus 2,000 km to Marry an AI Farm Girl With 300,000 Followers Business Anthropic Pledges 5 Percent of Itself to a Public Interest Fund That Grows With Its Valuation AI Agents NVIDIA Open-Sources ASPIRE, the Robot Framework That Turns Failed Grasps Into a Growing Skill Library Coding Assistants PostgreSQL 19 Slips to November 19 After LLM-Assisted Code Review Turned Up Forty Defect Reports News VAST Open-Sources EdenLive, a 5B Real-Time World Model That Tops 24 FPS at 720p on a Single RTX 5090 Business Nvidia Tells Samsung, SK Hynix and Micron to Build Its Custom HBM Base Dies on TSMC's 3nm Process Claude At Anthropic, Junior Share of Technical Staff Fell From 21 Percent to 14 Percent in a Year News Holyvolt Raises 45 Million Dollars at a 245 Million Valuation to Let AI Pick the Next Battery Chemistry

NVIDIA Open-Sources ASPIRE, the Robot Framework That Turns Failed Grasps Into a Growing Skill Library

AI Agents · By: AIReport Admin · 2 hours ago · 👁 7
NVIDIA Open-Sources ASPIRE, the Robot Framework That Turns Failed Grasps Into a Growing Skill Library

NVIDIA has released ASPIRE V1.0, an open-source agentic framework in which a coding agent diagnoses robot task failures from multimodal trajectories, rewrites the control program and banks verified fixes in a skill library of more than 90 skills covering 150-plus tasks. With GPT-6 Astra as the reasoning model, NVIDIA reports 99 percent success across 80 LIBERO-Pro tasks — company-reported figures that have not been independently verified.

NVIDIA has open-sourced ASPIRE V1.0, a robot agent framework built around a simple premise: treat the robot as a coding agent that can run, debug and improve its own control programs. The name stands for Agentic Skill Programming through Iterative Robot Exploration, and the release includes code, experiment configurations and the accumulated skill library.

The workflow addresses the most tedious part of robot programming. When a manipulation task fails — the cup gets knocked over, the gripper closes on nothing — an engineer traditionally digs through logs, adjusts parameters and re-runs. ASPIRE hands that loop to an agent. Every execution leaves a complete multimodal trajectory: what the camera recognized, which grasp candidates were computed, what path the arm planned, where contact and collision actually happened. On failure, the agent walks back through the trace, locates the broken step — wrong object identified, grasp offset, unreachable target, infeasible motion plan — rewrites the control program and sends the robot out again. Successful fixes are verified and stored.

The framework also explores multiple repair strategies in parallel: switching grippers, adjusting the robot's standoff distance, trying a different motion planner or approaching from another angle. Only methods that survive repeated execution enter the next stage, and that next stage is where the compounding happens — the skill library, which now holds more than 90 skills covering over 150 tasks, all downloadable by anyone running the framework.

The headline results are company-reported and should be read that way. According to Jim Fan, who leads the robot team at NVIDIA, when OpenAI released GPT-6 Astra the team swapped it in as ASPIRE's reasoning engine, and benchmarks jumped overnight: success on the 80 LIBERO-Pro tasks reached 99 percent, while representative task sets on Robosuite and BEHAVIOR-1K hit 100 percent. Fan called the result a zero-to-one tsunami for robotics and noted that Claude Opus 5.5 has since slightly overtaken Astra on the same stack. Earlier, with the previous model, ASPIRE managed 31 percent zero-shot on unseen LIBERO-Pro Long tasks against roughly 4 percent for prior methods.

Independent corroboration exists, with caveats. A separate October paper from NVIDIA, UCSD and UT Austin (Recova, arXiv 2610.01178) benchmarks against ASPIRE as the strongest baseline, reporting 71.7 percent average across six LIBERO-Pro settings — evidence the system is real and reproducible enough to be a reference point, though third parties have not replicated the post-Astra numbers.

What makes ASPIRE more than another benchmark demo is the economic structure it implies. The skill library is an appreciating asset: every failure the fleet debugs becomes reusable code, and when the underlying reasoning model improves, the entire library gets smarter overnight without new training. That is a different business shape from selling robots, and it is the part competitors should worry about.

Comments (0)

Log in to join the discussion

Log In

No comments yet