Sequoia Capital• Jan 6, 2026• 1:01:37Interview

Training General Robots for Any Task: Physical Intelligence’s Karol Hausman and Tobi Springenberg

Karol Hausman, Tobi Springenberg•Leaders at Physical Intelligence

Executive Summary

Physical Intelligence is building foundation models for robotics, aiming to create a general-purpose AI that can control any robot for any task, addressing the historical bottleneck of intelligence rather than hardware.
Their latest model, Pi Star 0.6, utilizes reinforcement learning (RL) from real-world physical experience, not simulation, to overcome performance plateaus seen with imitation learning alone.
This RL-based approach has dramatically improved performance, enabling a robot to serve coffee for 13 hours straight and increasing task throughput by over 2x, making commercial deployment feasible.
The company's core strategy is to deploy robots for economically valuable tasks, creating a data flywheel where real-world operations generate the vast, diverse data needed to further improve the foundation model's generalization and capabilities.

10 quotes

Concerns Raised

Generalization to completely new environments remains a significant, unsolved challenge.
The 'long tail' of real-world failures is a major obstacle to achieving the reliability needed for widespread deployment.
Simulation is insufficient for training manipulation tasks, making expensive and slower real-world data collection a necessity.

Opportunities Identified

Solving the 'intelligence bottleneck' could unlock the capabilities of existing and future robot hardware across countless applications.
Deploying robots for valuable tasks creates a powerful, self-sustaining data collection flywheel.
Foundation models for robotics show emergent capabilities, generalizing to new tasks and even new domains (e.g., driving, surgery) in unexpected ways.
Reinforcement learning from physical experience can dramatically increase task throughput and reliability, making commercial deployment viable.

Key Themes

Foundation Models for Robotics

Physical Intelligence is pioneering the development of general-purpose robotic foundation models. This approach contrasts with traditional, vertically-integrated robotics by focusing on creating a single, adaptable intelligence layer that can be applied to various hardware form factors and tasks.

This strategy aims to solve the core intelligence bottleneck that has historically limited robotics, potentially unlocking widespread adoption across industries without needing to build a bespoke software stack for every new application.

Reinforcement Learning in the Real World

The company has demonstrated a significant breakthrough by successfully applying reinforcement learning directly on physical robots. This allows the models to learn from their own experience, correct mistakes, and improve performance beyond what's possible with static demonstration datasets.

This 'real-world first' RL approach is critical for tackling the long tail of unexpected failures and environmental variations, a major hurdle for deploying reliable robots. It moves beyond the limitations of simulation, which often fails to capture the complexity of real-world physics and interactions for manipulation tasks.

The Data Flywheel Strategy

Unlike LLMs which can leverage vast internet text data, robotics lacks large-scale public datasets. Physical Intelligence's strategy is to create a data flywheel by deploying robots to perform useful, economically valuable tasks, which in turn generates more high-quality, real-world interaction data.

This creates a scalable and economically viable path to acquiring the massive, diverse datasets required for training truly generalist robot models. The cost of data collection becomes negative as the robots are performing valuable work.

End-to-End Learning vs. Classical Robotics

The speakers advocate for a paradigm shift from the classical approach of breaking robotics into separate perception, planning, and control modules. They champion a single, end-to-end neural network that learns directly from sensory input to motor commands, arguing this holistic approach is more effective and scalable.

This shift mirrors the evolution in other AI fields and suggests that trying to engineer separate components is less effective than allowing a large, general-purpose model to learn the entire process from vast amounts of data, leading to more robust and emergent capabilities.

Get started free

Topics

Robotics Foundation Models Physical Intelligence Reinforcement Learning (RL)Artificial Intelligence Machine Learning Data Collection Data Flywheel Embodied AI Vision-Language-Action Models (VLA)End-to-End Learning Generalization Real-World Deployment Sim-to-Real Robotic Manipulation

Processed May 4, 2026 yt-dlp + mlx-whisper + Gemini

You're reading a preview

Get started free →