←
Physical Intelligence126 posts
Physical Intelligence avatar

Physical Intelligence

@physical_int

Physical Intelligence (Pi), bringing AI into the physical world.

◷ Joined March 2024▧ 84 Photos and videos
Physical Intelligence avatar
Physical Intelligence@physical_int·Oct 13

Upload robot data, fine-tune our most capable models and connect your robots to cloud hosted inference through the API. We handle the infrastructure so you can focus on what’s unique to your use case. See what teams are building and join the beta at pi.website/blog/api-beta

Physical Intelligence avatar
Physical Intelligence@physical_int·Oct 13

Announcing the beta release of the Pi API. We’ve partnered with companies across diverse industries to adapt our models to their robots and applications, from harvesting strawberries to assembling servers. Now we’re opening access to more teams.

0:35
Physical Intelligence avatar
Physical Intelligence@physical_int·Oct 9

Excited to see our partners at @Ultraroboticsco announce their $50M Series A! Ultra is a great example of what's becoming possible when general-purpose robot foundation models meet real-world deployments!

Robotics has a distribution problem. Today, Ultra is announcing $62M in funding to solve it, including a $50M Series A led by @hiFramework. We believe that deployment is the bottleneck in robotics. It’s relatively easy to make a robot look impressive in a controlled environment. It’s much harder to put a robot into the open world, make it useful, keep it running, earn the trust of the people working alongside it, and do that over and over and over again. So that’s what we’re focusing on.

2:04
Shane avatar
⟳ Physical Intelligence retweeted
Shane@shinheeshanelee·Sep 8

4. Physical Intelligence X Deft: factory-proven foundation models Deft Robotics and Physical Intelligence are bringing general-purpose robot intelligence into real automotive factories. The task is line feeding: repeatedly picking up parts and placing them onto the production line in the correct position and orientation. It sounds simple. But parts shift around in the material bin as they’re picked. Each pick presents a different position and orientation, making fixed, rules-based automation impractical to deploy at-scale. Physical Intelligence’s robot foundation models, paired with Deft’s hardware and deployment stack, let the robot adapt to this variation. The robot locates each part, adjusts its grasp, and rotates it into position for placement—feeding the line without requiring every part to start in a predefined location. Across both deployments: 500 hours of operation and more than 10,000 parts processed. Every intervention surfaces the failures that matter, so the next model learns from real production failures. It's this iterative loop of deploying, intervening, retraining, and redeploying that makes the robot better over time — not just the cleanest demo. Robot intelligence becomes useful when it survives the factory.

1:35
Physical Intelligence avatar
Physical Intelligence@physical_int·Jun 18

The model also knows when not to trust itself: its uncertainty estimates track how far generations deviate from reality. The uncertainty accurately predicts when model generation degrades

Physical Intelligence avatar
Physical Intelligence@physical_int·Jun 18

Favorite finding: physical failures transfer. A policy that fumbles a grasp in the real world fumbles it in the world model too.

Physical Intelligence avatar
Physical Intelligence@physical_int·Jun 18

We found that the key to make this work is self-consistency: the model is trained to be self-consistent between a forward and inverse model (i.e., actions predicted with inverse dynamics lead to states consistent with ID input), consistent across different cameras, and consistent between training and test by halting episodes that go out of distribution. With and without self consistency comparison video

Physical Intelligence avatar
Physical Intelligence@physical_int·Jun 18

Across VLA checkpoints of varying performance on our table bussing benchmark, world model evals reproduce the real-world policy ranking. correlation plot: predicted vs ground truth success rate

Post attachment
Physical Intelligence avatar
Physical Intelligence@physical_int·Jun 18

Setup: • world model trained on our robot data generates what the cameras would see • VLA in the loop: sees generated frames, outputs actions, world model predicts what happens next

Physical Intelligence avatar
Physical Intelligence@physical_int·Jun 18

Why it matters: evals are critical to progress. As our models become more general and more capable, evaluating all the different things they can do takes longer and longer. The world model eval needs a couple hours of GPU time.

Physical Intelligence avatar
Physical Intelligence@physical_int·Jun 18

New work with @nvidia: evaluating robot policies entirely inside a world model. The policy acts, the model imagines the consequences, and the imagined evals predict real-world results. 🧵 real vs world-model rollout side by side📷

Physical Intelligence avatar
Physical Intelligence@physical_int·Apr 16

We are still discovering what π0.7 can do. It's fun to play with and the results so far have been quite surprising!

1:43
Physical Intelligence avatar
Physical Intelligence@physical_int·Apr 16

π0.7 handles diverse prompts that don't just say what to do, but also how to do it, including rich language and multimodal information, such as visual subgoal images. At test time, these images can be produced by a lightweight world model.

0:29