Robotics' New Era After Astra

An experiment published September 4 by RoboCurve found that OpenAI's GPT-6 Astra, a general-purpose model not trained as a robotics foundation model, controlled two robotic arms to pick up a block and place it in a bowl in 19 of 20 trials. A follow-up NVIDIA Astranomy study tested Astra across 120 simulated manipulation tasks, where its best configuration reached a 74.8% success rate, beating several specialized robotics models, though agentic runs cost about $36.40 per successful simulated episode with 18.56 seconds per inference call.

On September 4, RoboCurve published a simple experiment: OpenAI’s GPT-6 Astra controlled two robotic arms and was asked to pick up a block and place it into a bowl. It succeeded in 19 out of 20 trials . A month later, NVIDIA’s Astranomy study https://research.nvidia.com/labs/srl/projects/astranomy/ tested Astra across 120 simulated manipulation tasks. Its best configuration achieved a 74.8% success rate , significantly outperforming several specialized robotics models on the same benchmark. The surprising part? Astra was not designed as a robotics foundation model. A general-purpose model trained on reasoning, vision, coding, and tool use can now perform meaningful physical tasks without task-specific robot-policy fine-tuning. This IS the GPT-3 moment robotics has been waiting for. Before Astra: The Approach was Never Elegant Recommend this great primer by my friend Ella: There is an old joke in robotics: “The closer you are to embodiment, the further you are from intelligence.” For years, robotics progress was largely a game of incremental improvements and increasingly impressive demos. Unlike LLMs, robotics pre-training showed ~no scaling laws. There was little convergence on either model architecture or data strategy. To be honest, the industry was a bit uninvestable. - On models, the industry remained divided between VLA and WAM . - On data, companies pursued everything from internet video and first-person human demonstrations egocentric data to robot teleoperation and simulation. The common assumption was that robotics required substantial physical-world learning beyond what general-purpose AI models already knew. The bottleneck was data. To be fair, data scaling had begun to show results. NVIDIA’s EgoScale https://research.nvidia.com/labs/gear/egoscale/ demonstrated improvements using over 20,000 hours of human video, while Dyna-2 https://www.dyna.co/dyna-2 established a human-to-robot transfer scaling law using more than one million hours. But Astra suggests there may be another path. The obotics demo of the year did not come from a robotics company. What Actually Changed? Consider a robot picking up a glass. - Before Astra: A conventional VLA might learn from thousands of human or robot demonstrations across different objects, orientations, and grasping positions. The model learns to translate visual observations into executable actions and eventually generalizes to unfamiliar situations. - With Astra: The model can observe the glass through cameras, reason about how to approach it, generate commands through a robot-control interface, observe the result, and correct its next action. Of course, Astra is not yet a practical real-time controller. In NVIDIA’s Astranomy study, its agentic configuration cost approximately $36.40 per successful simulated episode , while direct action generation averaged 18.56 seconds per inference call . Those economics and latencies are prohibitive for many applications today. But both should improve through model optimization, distillation, and faster inference hardware. Several people I spoke with envision a future where a much smaller model, potentially around 100 billion parameters, runs locally on the robot. The Next Bottleneck: Deployment One consistent theme from my recent conversations with robotics researchers and engineers is how much work still goes into getting robots deployed. Even when the model is good enough, someone must understand the customer’s workflow, design data collection, build evaluations, diagnose failures, integrate hardware, and improve reliability. Much of that process still requires substantial researcher and engineering time. This creates a potential paradox: As frontier models become more capable, the robotics industry’s bottleneck may shift from model development to deployment engineering. And deployment is not just a cost center. It is also where proprietary physical data is created. Robots operating in the real world encounter hardware variation, unexpected objects, changing environments, and failures that are difficult to reproduce in the lab. 1. Implications on Specialized Robotics Models? Astra can already perform tasks historically assigned to System 2 and parts of System 1. It understands a scene, plans an approach, generates actions, and replans based on feedback. But it cannot replace System 0, yet. If a glass begins slipping from a robotic hand, the controller may need to adjust its grip within milliseconds. Waiting for a frontier model to reason through another action is not an option. This is why fast, specialized control remains necessary. My working hypothesis is that System 2 and much of System 1 eventually converge into general-purpose frontier models, while System 0 remains a smaller, faster, embodiment-specific controller. The open question is how far down the stack frontier models can reach. Could some of System 1 eventually be handled by standardized hardware APIs and control software rather than separately trained robotics foundation models? If so, the value of standalone robotics models could compress significantly. Coding offers an interesting counter argument: frontier models have become extremely capable programmers, yet specialized coding products have still found substantial markets. Robotics TAM could be large enough to support multiple winners. But specialized model companies will increasingly need to prove what they can do that general-purpose models cannot. 2. Implications on the Robotics Data Thesis? We do not know exactly what robotics-specific data Astra was trained on. Some say zero . My inclination is that Astra’s physical capabilities largely emerged from general-purpose multimodal pretraining rather than extensive robot-specific video-state-action trajectories. Astra may already understand what it means to pick up a glass. Showing it another thousand examples of people picking up similar glasses may teach it relatively little. My suspicion is that a substantial portion of today’s egocentric video is repetitive, relatively simple manipulation data. I would not be surprised if something approaching 80% of such footage offered limited incremental value to a sufficiently capable frontier model. What it still struggles with falls into two broad categories: - First, complex physical manipulation. Precision assembly, screwing small components, handling wiring harnesses, or manipulating deformable materials. - Second, physical information that vision alone cannot reliably provide. Tactile feedback, force, friction, pressure, and other sensor modalities. This shifts the discussion from how much data to what data, in what mixture, with what training recipe . Diversity also matters. Another underappreciated bottleneck is labeling quality. Ask ten operators to “pick up a spoon,” and they may execute the task ten different ways. Without more precise labels describing the intermediate actions and physical context, adding demonstrations can introduce ambiguity rather than useful information. Tactile sensing is another area I am watching closely. The industry is still experimenting with capacitive, visuotactile, electromagnetic, and other approaches. A model may understand that a glass is fragile. That does not mean it knows exactly how much force a particular robotic hand should apply. A growing data market: Almost every major AI lab is reportedly buying robotics data. Industry contacts describe million-hour procurement programs, highly specific requirements for data formats and collection hardware, and budgets of ~$1 billion at each leading labs. 3. Implications on Hardware? Astra is broadly positive for robotics hardware. If intelligence becomes easier to access, more companies can build capable robots without developing a frontier model themselves. That should accelerate hardware experimentation, deployment, and ultimately unit demand. Geopolitical: In July 2026, the FCC added certain foreign-produced advanced robotic devices to its Covered List, restricting authorization of new covered models without conditional approval. The domestic-end-product standard generally requires U.S. manufacturing and more than 65% domestic component cost through 2028 , rising to 75% in 2029. 3. Implications on Semis? A Potentially Enormous Edge-Compute Market Imagine a 100-billion-parameter model running on every humanoid robot. Even at 4-bit quantization, the model weights alone require roughly 50GB of memory, before accounting for runtime overhead and context. For semis investors, the questions are: Today robotics models sub 10B can run on Thor. How large must on-device models become? How aggressively can they be compressed? What’s NVDA’s roadmap? What is the optimal balance between cloud and edge inference? And will memory capacity or bandwidth ultimately constrain deployment? 4. Implications on Self-Driving? After watching general-purpose models transform coding and now begin to penetrate robotics, it is difficult not to ask the obvious next question: Could frontier models disrupt autonomous driving too? I think it is a real possibility. I would not call autonomous driving fully solved. Geographic generalization, long-tail safety, regulation, and operational economics remain important challenges. But commercial robotaxi is increasingly becoming a deployment and scaling problem. One possible outcome is the Apple/Android split in which leading AI models become autonomous-driving platforms licensed to multiple OEMs. That would have significant implications for the competitive advantages of TSLA 0.00%↑ https://substack.com/search/%24TSLA and traditional automakers. Final Thoughts I’m 100x more bullish on robotics as an industry than ever. I’m increasingly skeptical of the long-term value of standalone robotics-specific models . I keep coming back to a simple intuition: perhaps elegance matters. For years, robotics has felt incredibly fragmented. Every company has its own approach to data collection, model architecture, and hardware design. There isn’t even a consensus on whether robots should be humanoid, let alone how a humanoid should be designed. Multiply all these possibilities together, and you get an industry pursuing countless different paths, with few standardized benchmarks to tell us which one will scale and none of them showed any real scaling laws . Then Astra came along with a remarkably elegant idea: What if we don’t need to reinvent general intelligence for robotics? What if we can simply bring it into the physical world? Astra hasn’t solved robotics, but for the first time, the path forward feels much clearer. Maybe there’s something to be said for trusting that intuition. I also do believe, the largest TAM will yield some of the highest investment returns. Go robotics