Physical Intelligence Launches Pi API Beta for Hosted Robot Foundation Models
Developers gain hosted fine-tuning and inference access to iterate rapidly across diverse robotic platforms.
1Pi API Beta Launch
Physical Intelligence officially unveiled its Pi API beta on October 13, 2026, granting external developers access to its advanced robotic foundation models. Access operates through a selective waitlist, requiring applicants to outline specific use cases before testing hosted fine-tuning pipelines and cloud inference directly on their own robotic hardware setups.
2Hosted Workflow Mechanics
The API simplifies deployment by taking over complex infrastructure demands. Developers upload task demonstration data, trigger fine-tuning jobs on hardware managed entirely by Physical Intelligence, and access the resulting policy through a hosted gateway. Real-world evaluation data from physical robots then directly informs subsequent fine-tuning iterations to boost task precision.
3Open Source Model Lineup
The open-source openpi repository offers three core architectures: π0, an action-flow model; π0-FAST, an autoregressive variant using the FAST action tokenizer; and π0.5, optimized via knowledge insulation. Pre-trained on over 10,000 hours of robot telemetry, these Apache-2.0 licensed models offer PyTorch support and full DROID dataset compatibility.
4Local Hardware Constraints
Executing openpi models locally imposes steep hardware barriers. Basic inference demands an NVIDIA GPU with at least 8 GB VRAM, such as an RTX 4090, while full fine-tuning requires 70 GB memory on A100 or H100 systems. The API eliminates these compute constraints via remote websocket action streaming.
5Heavy Industry Validation
To evaluate architectural adaptability across disparate form factors, the company partnered with industrial teams. Actor deployed an autonomous wheel loader in Utah to scoop materials and fill dump trucks without human operators. Other industry pilot partners include AgRobotics, CloudChef, Ultra, Weave, and commercial dairy operations managing unstructured environments.
6Iterative Learning Pattern
Pilots revealed a uniform operational bottleneck across diverse domains, whether resolving twisted garments or handling spilled culinary ingredients. Teams collect physical demonstrations, execute fine-tuning, benchmark failure modes on robots, and harvest edge cases for targeted data collection. The API was specifically designed to accelerate this cyclic loop independently.
7Rapid Assembly Iteration
Early access partner Almond utilized the Pi API to program its Axol dual-arm robot. Seeking to enable the robot to assemble its own components, Almond trained Axol to deburr mechanical spacers after only four days of iterative fine-tuning, validating the platform's speed in achieving tangible industrial manipulation capabilities.
8Global Autonomous Deployments
Early API deployments expanded across Japan, Switzerland, and the United States, engaging startups like Formic, Chef Robotics, and Frontier Machines. While remote websocket inference validated cross-platform portability, trials highlighted edge infrastructure vulnerabilities such as live network latency, defining core operational focus areas for upcoming platform updates and enterprise recruitment.