{"slug": "nvidia-s-open-source-aspire-system-taught-a-robot-to-solve-a-rubik-s-cube", "title": "Nvidia's open-source ASPIRE system taught a robot to solve a Rubik's Cube", "summary": "Nvidia's GEAR lab released version 1.0 of ASPIRE on GitHub this month under an Apache 2.0 license, reporting 99% success across 80 tasks in the LIBERO-Pro benchmark and a robot arm solving a Rubik's Cube without task-specific training. ASPIRE is a code-as-policy system in which a large language model writes Python control code, observes execution, and rewrites failures into a stored skill library, replacing the reinforcement-learning approach used by OpenAI's 2019 Dactyl hand. The project's own benchmark showed swapping the underlying model from GPT-5.6 Luna to Claude Opus 5.5 raised success on ten LIBERO-Pro tasks from 15% to 99% on identical hardware, while the Rubik's Cube run used GPT-6 Astra.", "body_md": "*99% success across 80 manipulation tasks, no trial-and-error training, and a Rubik's Cube solved along the way. Nvidia open-sourced the system behind it this month.*\n\nNvidia's GEAR lab released version 1.0 of ASPIRE on GitHub this month, and the headline demo is a robot arm solving a Rubik's Cube without ever being trained specifically to do it. According to the project's NVlabs repository, the October update pushed the system to 99% success across 80 tasks in the LIBERO-Pro benchmark, spanning goal, object, spatial, and long-horizon manipulation challenges, plus selected tasks from Robosuite and BEHAVIOR-1K. The code is released under an Apache 2.0 license, meaning anyone can download, modify, and run it.\n\nWhat makes ASPIRE different from the robotics demos Nvidia has shown before is the method, not just the trick. Most dexterous-manipulation systems, including OpenAI's well-known 2019 Rubik's Cube hand, Dactyl, were built on reinforcement learning: a policy trained for months in simulation through millions of trial-and-error attempts, then transferred to physical hardware. ASPIRE skips that entirely. It's a code-as-policy system. A large language model writes actual Python control code for the robot, watches what happens when that code runs, and rewrites it when it fails. Validated fixes get stored in what the project calls a skill library, so the next task doesn't start from zero.\n\nASPIRE first appeared as a research paper on June 30, posted to arXiv by a team including Nvidia GEAR co-founders Jim Fan and Yuke Zhu, alongside collaborators from the University of Michigan, the University of Illinois Urbana-Champaign, UC Berkeley, and Carnegie Mellon. That original version reported up to a 77-point improvement over the strongest prior baseline on LIBERO-Pro under perturbation, a 72% success rate on a Robosuite bimanual handover task, and 31% zero-shot success on LIBERO-Pro's hardest long-horizon tasks, against roughly 4% for earlier methods. There was no Rubik's Cube in that version. That capability, along with the version 1.0 label, arrived with the October release.\n\nThe project's own benchmarking makes an uncomfortable point for anyone assuming better robots just need better hardware. On ten LIBERO-Pro tasks, swapping the underlying language model from GPT-5.6 Luna, OpenAI's cheap, fast model, to Claude Opus 5.5 took success from 15% to 99%, using the identical robot and the identical code-execution engine. The real-world Rubik's Cube run was driven by GPT-6 Astra, OpenAI's top-tier model released this fall, while a separate table-bussing demo ran on GPT-5.5. Nvidia's own documentation even plots what it calls a cost-success frontier, acknowledging flatly that the priciest model isn't always the one worth paying for. The robot arm, in other words, was never the limiting factor. The reasoning engine behind it was.\n\n[Nvidia is in talks to buy the open-source AI lab it just backed](https://startupfortune.com/nvidia-is-in-talks-to-buy-the-open-source-ai-lab-it-just-backed/)\n\nReflection AI's valuation has tripled to $25 billion in under a year, and Nvidia, already an $800 million investor, is weighing options from an acqui-hire to a bigger equity stake, Bloomberg and the Financial Times report. - [Nvidia acquisition talks with Reflection AI lab](https://startupfortune.com/nvidia-is-in-talks-to-buy-the-open-source-ai-lab-it-just-backed/) - [how Nvidia is buying open source AI startups](https://startupfortune.com/nvidia-is-in-talks-to-buy-the-open-source-ai-lab-it-just-backed/)\n\nThat's a sharper claim than it sounds. For two years, the \"physical AI\" pitch from Nvidia CEO Jensen Huang has leaned on the idea that robotics is the next trillion-dollar platform after data-center GPUs, with Nvidia selling the simulation and training stack underneath it. A TechCrunch piece in January framed Nvidia's robotics ambitions as wanting to be \"the Android of generalist robotics,\" supplying the operating layer other humanoid and arm makers build on top of. ASPIRE fits that strategy precisely: open-source it, let outside labs and startups build on Nvidia's skill-library architecture, and the actual intelligence driving it can come from any frontier LLM, OpenAI's or Anthropic's included.\n\nNvidia isn't hiding the rough edges, either. The GitHub documentation includes a blunt warning that would never make it into a polished product announcement: \"ASPIRE executes language-model-generated Python with full import access. Trial processes and watchdogs are not a security sandbox.\" That's a real caveat for anyone planning to run this outside a research lab, since it means a misbehaving model isn't just risking a failed cube turn, it's running arbitrary code with real permissions.\n\nFrankly, that tradeoff is the whole story here. Chatbots proved language models could reason about text. ASPIRE is a demonstration, open and reproducible, that the same models can reason about a robot's control loop well enough to debug it themselves, face turn by face turn, until a toy puzzle gets solved. It isn't a finished product and Nvidia isn't claiming the \"ChatGPT moment\" for robotics has arrived. But it's a concrete, downloadable result, not a demo reel, and that puts real pressure on every closed robotics stack claiming to be ahead.\n\n**Also read:** [Self Labs Launches Agent Vault as Banks Lock AI Agents Out of Accounts](https://startupfortune.com/self-labs-launches-agent-vault-as-banks-lock-ai-agents-out-of-accounts/) • [Jamie Dimon Bars Outside AI Agents From JPMorgan Customer Accounts](https://startupfortune.com/jamie-dimon-bars-outside-ai-agents-from-jpmorgan-customer-accounts/) • [PrismML's ternary AI model runs a 27B brain in under 6 gigabytes](https://startupfortune.com/prismmls-ternary-ai-model-runs-a-27b-brain-in-under-6-gigabytes/)\n\n*This article is posted in [AI News](https://startupfortune.com/category/ai/), check it out for more related stories.*\n\n## Join the discussion\n\n[Open in the community →](https://startupfortune.com/community/)\n\nAlmost there. Sign in and your reply posts straight away.", "url": "https://wpnews.pro/news/nvidia-s-open-source-aspire-system-taught-a-robot-to-solve-a-rubik-s-cube", "canonical_source": "https://startupfortune.com/nvidias-open-source-aspire-system-taught-a-robot-to-solve-a-rubiks-cube/", "published_at": "2026-10-16 18:38:30+00:00", "updated_at": "2026-10-16 18:53:20.994126+00:00", "lang": "en", "topics": ["robotics", "artificial-intelligence", "large-language-models", "ai-research", "ai-agents"], "entities": ["Nvidia", "Nvidia GEAR", "ASPIRE", "Jim Fan", "Yuke Zhu", "OpenAI", "Dactyl", "LIBERO-Pro"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/nvidia-s-open-source-aspire-system-taught-a-robot-to-solve-a-rubik-s-cube", "markdown": "https://wpnews.pro/news/nvidia-s-open-source-aspire-system-taught-a-robot-to-solve-a-rubik-s-cube.md", "text": "https://wpnews.pro/news/nvidia-s-open-source-aspire-system-taught-a-robot-to-solve-a-rubik-s-cube.txt", "jsonld": "https://wpnews.pro/news/nvidia-s-open-source-aspire-system-taught-a-robot-to-solve-a-rubik-s-cube.jsonld"}}