Untrained Tasks Now Possible via π0.7 Robotic Brain

Physical Intelligence released π0.7, a steerable robotic foundation model capable of untrained task execution. The April 2026 release demonstrates compositional generalization, allowing robots to combine fragmented knowledge from disparate training data to solve novel physical problems. For developers building robotic control systems, this shifts the production paradigm from training specialized models for individual tasks to prompting a single multi-robot foundation model.

Architecture and Training Logic

π0.7 builds on the company’s previous Vision-Language-Action (VLA) architectures. The model ingests Internet-scale vision-language pretraining data alongside specialized multi-robot datasets. It processes this multimodal input and directly outputs low-level motor commands.

The core capability relies on compositional generalization rather than rote memorization. During internal testing, the model successfully operated an air fryer despite having only two relevant training episodes in its entire dataset. One episode showed a robot closing a door. The other showed a robot placing a bottle inside an object. The model synthesized these isolated behaviors to complete the novel air fryer task. If you build multi-agent systems or physical robotic controllers, this capability drastically reduces the burden of gathering exhaustive task-specific demonstration data.

Performance Benchmarks

In general-purpose internal testing, π0.7 matched the performance of previous specialist models. Specialist models are traditionally trained to execute exactly one function. π0.7 handled diverse physical operations without task-specific fine-tuning.

Demonstrated tasks include:

• Making coffee
• Folding laundry
• Assembling boxes
• Cleaning entire kitchens and bedrooms using mobile manipulators

UC Berkeley Professor and Physical Intelligence co-founder Sergey Levine noted that the model’s capabilities are scaling more than linearly relative to the amount of data collected. Validating these physical models across varying environments requires rigorous methodology, similar to how you evaluate and test AI agents in software interfaces.

Hardware Decoupling and System Upgrades

Physical Intelligence positions π0.7 as a universal operating system for robotics. The underlying architecture is designed to decouple the artificial brain from specific hardware bodies.

This release follows several rapid architectural iterations. The November 2025 π0.6 model utilized reinforcement learning to improve general success rates. In March 2026, the company introduced Multi-Scale Embodied Memory (MEM), allowing models to manage continuous tasks lasting longer than ten minutes. They simultaneously released an RL Token method to extract reinforcement learning tokens directly from VLA models for fast online learning. Implementing similar long-horizon task execution in your own software applications often requires dedicated architectural work to add memory to AI agents.

These technical deployments parallel a massive influx of capital. Reports from late March 2026 indicate Physical Intelligence is raising $1 billion at an $11 billion valuation, potentially led by Founders Fund and Lightspeed Venture Partners. This follows a $600 million round at a $5.6 billion valuation in late 2025.

If you deploy robotic hardware, you should evaluate steerable foundation models against your existing specialist stack. The ability to generate reliable low-level motor commands from a generalized VLA model means you can prioritize hardware integration and prompt orchestration over collecting thousands of manual training demonstrations.