By Sarah Xiao 
NVIDIA’s GEAR lab open-sourced ASPIRE v1.0 on 16 October, and the number that travelled is 99%. It is the right number to be suspicious of, and the right one to understand.
ASPIRE — Agentic Skill Programming through Iterative Robot Exploration — does not train a policy. It hands the robot a language model that writes ordinary Python control code, watches the robot run it, reads the trace when it fails, patches the code, and tries again. Fixes that survive get written into a skill library as retrievable entries. The library is the product; the model is a swappable component.
The published results, all from the project’s own benchmarking:
- 99% success across 80 tasks on LIBERO-Pro, plus selected tasks from Robosuite and BEHAVIOR-1K at 100%.
- 31% zero-shot on LIBERO-Pro’s long-horizon unseen tasks, against about 4% for prior methods.
- Skill library of 90+ skills, 150+ tasks, 18 categories — grasping, navigation, object search, perception, bimanual work, failure recovery.
- Code on GitHub under Apache 2.0; paper submitted to arXiv on 30 June, with collaborators from Michigan, UIUC, UC Berkeley and CMU.
The number that actually matters is 15%
Forget 99% for a moment. The finding worth acting on is a different one, and it comes from a controlled comparison the team ran.
On ten LIBERO-Pro tasks, they swapped only the underlying language model — from a cheap, fast model to a frontier one. Success went from 15% to 99%. Identical robot. Identical code-execution engine. Identical skill library. Nothing retrained.
That single number reframes what “robot capability” means. If a factor-of-six improvement comes from a component that has nothing to do with the robot, then the robot was never the limiting factor in that benchmark. The reasoning engine writing the code was.
There is a second half to this that is more useful than the headline: the team publishes what it calls a cost-success frontier, and notes plainly that the most expensive model is not always the one worth paying for. In the project’s own reporting, Claude Opus 5.5 slightly edges out the top-tier Astra model on some robot benchmarks, while Sonnet 5.5 sits on the Pareto frontier at roughly $12 per task. Anybody planning a deployment should be reading that table, not the 99%.
Jim Fan’s framing — a “0 to 1 tsunami”, and a “Physical Bitter Lesson” — is that capabilities learned in software (reasoning, coding, tool use, long-horizon planning) now transfer into the physical world through frameworks like this. Whether or not you accept the framing, the architecture implies something concrete for buyers: the robot’s capability curve is now tied to a release schedule you do not control.
What the framework actually is
Structurally it is three parts, and the analogy to an agent harness is accurate.
- A closed-loop execution engine. Every perception call, grasp candidate, motion plan and collision result leaves a multimodal trace with visual evidence. When a task fails, the agent walks back through it: was the object recognised, was the grasp offset, could the arm physically reach, did the planner produce a traversable path. This is the difference between a system that reports “failed” and one that can be debugged.
- A skill library of validated patches. This is the part that persists. Grasp-point heuristics, perception prompts, planner selection, recovery procedures — stored as retrievable units rather than left in a researcher’s notebook. NVIDIA’s stated motivation is the old problem: everything learned in one task used to be discarded when the task ended, so the next task started from zero.
- Evolutionary search over candidate programs. The system generates several control programs in parallel and lets their failure traces inform each other, so it does not grind away on a single wrong approach. When the arm keeps failing to lift a can from the floor, it tries the other hand, changes the standoff distance, swaps in a heavier planner, approaches from a different direction.
The reported proof points: a robot arm learning to solve a Rubik’s cube online, rewriting its own control code between attempts, and a table-bussing demo where skills learned in simulation transferred to a different real machine — cutting the reasoning tokens needed for a first successful commissioning.
Four things a buyer should not take from this
99% is a simulation benchmark. The real-machine evidence in the announcement is a small number of demos — a cube, a table. The 80-task figure is LIBERO-Pro, in simulation, scored by the project. It is a strong result on its own terms and it is not a claim about your cell.
The security caveat is not a footnote. NVIDIA’s own repository warns that ASPIRE executes language-model-generated Python with full import access, and that the trial processes and watchdogs are not a security sandbox. Read that carefully: a misbehaving model is not merely risking a bad grasp, it is running arbitrary code with real permissions on the machine that controls the robot. That is acceptable in a research lab and categorically unacceptable on a production line without containerisation, allow-lists and a human in the loop.
Code-as-policy changes your validation obligations. If the control program is generated rather than authored, then the artefact you validate is not a fixed program — it is a program generator plus a skill library. Your quality system needs an answer to: which version of the library ran, on which model, with what trace, on this unit. PPAP does not yet have a field for that.
The skill library is the lock-in, and it is open. This is the strategically interesting part. NVIDIA is not selling robots; it is publishing the layer other people’s robots run on. If the skill library becomes the durable asset, then whoever hosts it sits where Android sits. The Apache 2.0 licence is what makes that plausible — and it is also what makes it hard to build a moat on.
Why this connects to the rest of the week
Three other things landed in the same window, and they are the same argument from different directions.
Amazon’s Whole-Body Dexterity (with UCSD and Michigan) shows the other half of the same shift: instead of better reasoning, better embodiment. Treating arms and torso as contact surfaces took a real Unitree G1 from 4% to 90% on whole-body manipulation tasks. That is not a smarter model — it is a better formulation of what counts as manipulation.
Returns logistics is where both ideas meet a paying customer: Sereact’s Cortex model is deployed at Zalando and CEVA facilities precisely because a returned garment cannot be pre-programmed for, and no per-SKU training is affordable.
Capital followed: RMB 1.7bn to an embodied-compute SoC maker, and over RMB 100m to a five-month-old joint-module startup. Money is moving toward whichever layer is believed to be scarce — compute, joints, or skills. This week it went to all three.
The honest summary
The durable contribution of ASPIRE is not a benchmark score. It is the demonstration that a robot’s capability can improve without touching the robot, and that the improvement can be stored as retrievable, transferable text rather than as weights.
For a factory, that is a new kind of asset with a new kind of risk: an asset you can audit and read, that improves on someone else’s release schedule, and that executes code you did not write.
Related reading
- Whole-body dexterity: stop using just hands — the other 2026 route to better manipulation.
- Returns are harder than picking — where “no per-SKU training” becomes a business requirement.
- 200 hours to 1 hour: the gearbox cell that writes its own insertion — the industrial version of the same shift, already in series production.
- The invisible hours: what robot commissioning really costs — the labour this framework is aimed at.
- Dream Labs: robot intelligence at $3 an hour — the commercial model for selling capability instead of capability-building.
Sources: ASPIRE v1.0 release (16 October 2026) via NVIDIA GEAR / NVlabs, Apache 2.0; the 99%-on-80-tasks result, the 31% zero-shot figure, the 90+/150+/18 skill-library counts, the 15%-to-99% model-swap comparison and the ~$12-per-task Sonnet 5.5 datapoint are from the project’s own reporting as reported by Startup Fortune, TechMoon and AI Insider; the arXiv paper (30 June submission) reported up to 77 points over baseline on LIBERO-Pro under perturbation. Whole-Body Dexterity figures are from arXiv 2610.22184 (Amazon, UCSD, Michigan). The validation, security and lock-in readings are Latest Robot News editorial.