Robot-Use Agents: Why General-Purpose Models May Win in Robotics
Robot-Use Agents: Why General-Purpose Models May Win in Robotics
Podcast29 min 49 sec
Listen to Episode
Note: AI-generated summary based on third-party content. Not financial advice. Read more.
Quick Insights

Treat robotics AI as a long-term theme to monitor, not a proven near-term investment: look for lower-latency control, reliable performance on unfamiliar tasks, and economically repeatable deployments. The discussion provides no specific publicly traded stock, price target, or validated company opportunity.

Detailed Analysis

Robotics AI and Robot-Use Agents

  • The discussion is bullish on the long-term potential of general-purpose AI models to control different kinds of robots, rather than relying only on robotics-specific models.
  • The proposed advantage is that general-purpose models can draw on broad training data—such as coding, images, and computer-use data—and use tools or write programs to handle physical tasks.
  • The speakers said some frontier labs and robotics-model companies expect general-purpose robots within two years or earlier. They described the goal as robots handling natural-language instructions for tasks a competent teenager could do by hand. This is a forecast, not a guaranteed timeline.
  • Risks and constraints mentioned: Model latency can make robot control too slow to be economically useful; robotics data is a bottleneck; and society may be unprepared for rapid capability improvements.
  • Actionable takeaway: Treat robotics AI as an emerging theme to monitor, not as a proven near-term investment thesis. Watch for evidence of lower-latency control, reliable performance on unfamiliar tasks, and repeatable deployments that make economic sense.

Waddle Labs / Model Labs (Private Startup)

  • The transcript describes a startup building language models and a harness—the tools and software around a model—to control robots, while collecting data to improve performance. The company is referred to as Waddle Labs and Model Labs in different parts of the transcript.
  • The speakers’ approach is to use a general-purpose model for flexible reasoning, then package learned tasks into faster, reusable skills or programs.
  • Risk mentioned: Keeping a large model in the control loop at every step can be too slow for practical use.

Takeaways

  • The company represents a potential private-market opportunity in robot-control software and data, but the transcript provides no valuation, funding details, or commercial results.
  • For this type of business, monitor whether its harness can turn successful demonstrations into fast, reliable skills that work across tasks and robot types.

RoboCurve (Private Startup)

  • RoboCurve describes itself as an evaluation company for physical AI, measuring the performance of different models and approaches across robots such as hands, grippers, arms, humanoids, and quadrupeds.
  • The discussion highlights evaluation as part of the broader robotics ecosystem: investors and developers need ways to compare model performance across tasks and embodiments.

Takeaways

  • Evaluation tools may benefit as more AI models are applied to robotics, but the transcript does not provide revenue, customers, or other evidence of commercial traction.
  • Follow whether its evaluations become useful, repeatable benchmarks for comparing real-world robot performance.

OpenAI and General-Purpose AI Models

  • OpenAI is mentioned in a discussion of advances in general-purpose models, including improved vision, tool use, and spatial reasoning. The speakers suggest that broader data—such as computer-use and CAD-related data—may help models perform physical tasks.
  • The discussion argues that a strong general-purpose model could potentially outperform a weaker model built only for robotics, if their representations of the world overlap sufficiently.
  • No specific stock, price target, or investment recommendation was given for OpenAI or any model discussed.

Takeaways

  • The investment theme is the possible extension of general-purpose AI capabilities into robotics—not a specific model or company recommendation.
  • A key question is whether model improvements translate into dependable robot performance outside curated demonstrations.

Google DeepMind and RT-2 Research

  • Google DeepMind’s RT-2 is discussed as an early example of using a pretrained language model, trained on web text and images, to control robots.
  • The speakers say its broader pretraining helped compared with training a robotics model without that foundation.
  • This is a research reference; the transcript does not provide an investment view on Google or its stock.

Takeaways

  • The discussion supports the theme that data and pretrained capabilities may matter as much as a model’s architecture in robotics.
  • The transcript offers no specific recommendation or price target for Google.

Tesla

  • Tesla is mentioned as a hypothetical example of a company with extensive self-driving data. A speaker argues that in a data-rich setting, learning from examples in context alone would not be an adequate substitute for training and updating the model.
  • No assessment of Tesla’s business, stock, or robotics prospects is provided.

Takeaways

  • Tesla’s mention is illustrative, not a bullish or bearish investment thesis.
  • The relevant broader point is that data-rich applications may need different learning methods from low-data robot deployments.

Figure

  • Figure is mentioned hypothetically as an example of a robotics company that would not rely on in-context learning alone if it had access to abundant data.
  • The transcript gives no specific company update, commercial details, or investment recommendation.

Takeaways

  • Figure is a private-company reference, not a publicly traded security identified in the discussion.
  • The point to monitor across robotics companies is how they combine accumulated data with faster, reusable control systems.
Ask about this postAnswers are grounded in this post's content.
Episode Description
One of the biggest surprises in AI over the last few years has been how well coding agents generalize beyond software. In a recent essay, MIT professor Philip Isola argued that we may be entering the era of robot-use agents: general-purpose models that can control different robots, write policies, and learn new physical tasks with little or no robot-specific training.In this episode of Decoded, we're joined by the founders of Waddle Labs and RoboCurve, two of the startups whose work helped drive this realization. They're working at the frontier of using general-purpose models to control robots, and their recent demos helped inspire the growing conversation around robot-use agents. Together, we dig into the research behind that idea, from code-as-policies and vision-language-action models to the harnesses and evals needed to make these systems work in the real world.
About Y Combinator Startup Podcast
Y Combinator Startup Podcast

Y Combinator Startup Podcast

By Y Combinator

We help founders make something people want.