NVIDIA open-sources ASPIRE, which cut Rubik's Cube turn time from 17 to 3.1 minutes in three episodes

Guanzhi Wang and Jim Fan of NVIDIA's GEAR lab announced on Oct 16 that ASPIRE v1.0 is open source on GitHub. The lab first shipped ASPIRE in June 2026; the real-robot results are new.

What the agent learns is stored as code and text rather than in model weights, and the gain shows up as time and token cost. Across three real-robot episodes, execution time per Rubik's Cube face turn fell from 17 minutes to 3.1 minutes, and tokens per turn dropped 82% as ASPIRE accumulated skills.

ASPIRE writes and refines code-as-policy robot control programs from execution feedback, inspecting rollout traces, repairing programs and saving reusable skills, the project page says. A closed-loop engine records observations, inputs and outputs for each perception, planning and control call, and a skill library distills validated fixes into guidance retrieved in context. The README calls the process continual learning, where "training" is skill refinement and the "trained model" is a repo of sensorimotor skills. The June release included table-bussing sim-to-real demos.

A symbolic solver writes the cube moves; the agent does the rest

On the real robot, ASPIRE with GPT-6 Astra solved a Rubik's Cube on a bimanual YAM robot with soft grippers, executing a 16-move solution and coordinating both arms to grasp, rotate and manipulate the cube, Wang said. "No teleoperation demos or simulation training required," he wrote. The operator runbook says the agent needs "an available symbolic cube solver to compute and validate a solution for the current scramble"; its own work is perception, grasping and turning. The project page says the per-turn figures are amortized over the full episode, including failed attempts and recovery; at those rates a 16-move solve works out to roughly 50 minutes at the third episode's pace (16 x 3.1) versus about 4.5 hours at the first (16 x 17), an estimate rather than a measured solve time.

The same 10 LIBERO-Pro tasks score 15% or 99%, depending on the model underneath

The team tested ASPIRE across GPT-6 Astra, Claude Opus 5.5, Claude Sonnet 5.5, Claude Fable 5.1 and open-weight models including Kimi K3 and Qwen 3.8 Max, holding the environment, robot programming API and evaluation protocol fixed. On 10 LIBERO-Pro tasks, success ranges from 15% with GPT-6 Luna to 99% with Claude Opus 5.5, and the strongest models are not the most expensive; Sonnet 5.5 sits on the cost-versus-capability frontier at $12 per task. Fan wrote that the day Astra arrived, the lab switched backends and saturated the benchmarks to about 99% overnight: "a rising frontier tide lifts all boats, but Astra was a 0->1 tsunami on many tasks." Simulated results are 99% on LIBERO-Pro's 80 tasks and 100% on 7 representative Robosuite tasks and 2 representative BEHAVIOR-1K tasks.

The three new real-robot recipes, including putting a pill into a bottle and screwing the cap on, are distilled from demo sessions; the README calls them "not a new real-robot validation." ASPIRE learns on a small debug split and is scored on disjoint held-out seeds. ASPIRE relies on a frozen frontier model, and the researchers say they have not verified that smaller or weaker models can sustain the debugging loop.