Jan 6, 2026, 8:00 AM ETInterview

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

Physical Intelligence is building robotic foundation models designed to control any robot form factor and execute any task.
The company views the primary bottleneck in robotics as "intelligence" rather than hardware capabilities.
The strategy is to avoid vertical integration for specific tasks, aiming instead for a general-purpose intelligence layer applicable to homes, industry, and specialized fields.
Founders Carol and Toby argue that classical robotics architectures (perception, planning, control) are fundamentally flawed because the interfaces between these sub-problems break down in complex real-world scenarios.
The company's approach is end-to-end learning, training a single model to map raw sensory inputs directly to robot actions.

The current model architecture is analogous to Vision-Language Models (VLMs), processing image and text inputs to generate text and action outputs.
The core model is a transformer with up to a billion parameters, pre-trained on internet data and fine-tuned on proprietary robotics data.
An "action expert" module was added to the VLM backbone to translate visual and linguistic understanding into specific robot control commands.
The system relies on human demonstration data (teleoperation) for initial policy training before reinforcement learning takes over.
The architecture is expected to evolve over the next 5–6 years, potentially moving away from current VLM-based backbones as physical reasoning capabilities mature.
Reasoning capabilities are emerging naturally, with the model decomposing high-level tasks into sub-tasks (e.g., "drive to counter" then "pick up glass") and predicting action horizons of approximately 50 time steps.

Physical Intelligence faces a unique data scarcity problem: unlike LLMs, there is no massive, free internet dataset for robot actions.
The company is currently in a "bootstrap phase" utilizing diverse data sources (simulation, teleoperation, internet video) to reach a deployable threshold.
Founders anticipate that the "deployment phase" will generate vastly more data than any bootstrap effort, with robots collecting data autonomously while performing economically valuable tasks.
Data quality and diversity are prioritized over sheer quantity; simply collecting more of the same data leads to performance plateaus.
The company is exploring "value functions" to identify successful vs. unsuccessful trajectories earlier in a sequence, rather than waiting for a final reward signal, improving data efficiency.

Pi-Star 0.6 introduces "RL from experience," allowing the robot to improve its policy by learning from its own interactions and human corrections during deployment.
This shift from pure imitation learning to RL from real-world experience enabled the model to escape the performance plateau seen in previous releases.
All RL training for Pi-Star 0.6 was conducted on real-world physical systems, not simulation, to capture rare failure modes (e.g., non-perforated cardboard boxes sticking together) that simulators missed.
The model achieved over a 2x increase in throughput across three specific tasks: building chocolate boxes, making coffee, and folding laundry.
In live testing, the robot successfully served coffee for 13 hours straight and folded laundry for 4 hours, demonstrating unprecedented reliability and failure recovery.
Human corrections (e.g., adjusting the force used to tamp coffee) were fed back into the model, resulting in rapid skill acquisition within 30–50 episodes.
While current RL improvements are task-specific (e.g., better tamping does not immediately improve box building), the accumulated data from diverse tasks is expected to drive future generalization.

The company has achieved "zero-shot" generalization to new environments, allowing robots to operate in homes they have never visited without specific retraining.
Founders state they are now "fairly close" to commercial deployment, having started actual deployments two months prior to the interview.
The "aperture" for deployment is wider than initially expected, with the technology being applied to diverse fields including driving, surgery, and agriculture via open-source releases.
However, full household deployment is not yet deemed safe due to privacy, safety, and the "long tail" of rare failure scenarios.
The company expects to deploy initially in controlled environments or non-critical tasks (e.g., 95% reliability is acceptable for laundry folding) before expanding to higher-stakes applications.
The founders reject the "best of both worlds" approach of mixing hardcoded physics rules with learning, arguing that end-to-end learning from experience is superior for adapting to complex physical realities.

Physical intelligence is viewed as a testbed for improving reasoning in LLMs, suggesting that grounding language models in physical actions and video predictions could solve the "counterfactual" and "credit assignment" problems currently limiting text-only AI.
The success of general-purpose models suggests that the AI field has moved past the era of breaking problems into isolated sub-components, with end-to-end learning proving more effective than modular robotics stacks.
The founders express that the fact these loosely brain-inspired, general-purpose algorithms work so well across vision, language, sound, and robotics is "mind-blowing" and fundamentally changes the trajectory of the industry.