Image: physicalintelligence.company
VLA that learns from its own experience via RL, demonstrated making espresso, folding laundry and assembling boxes for hours.
Key specs
| Method | Recap RL with value function on top of π0.6 VLA |
|---|
| Result | >2x throughput vs base model on hardest tasks |
|---|
Technical details
π*0.6 is Physical Intelligence's π0.6 VLA further trained with Recap (RL with Experience & Corrections via Advantage-conditioned Policies). Recap trains a value function on robot experience, uses changes in value as advantages, and conditions the VLA on advantage so it can be prompted to take high-advantage actions while still training on all data. Training stages are offline-RL pretraining, demonstration fine-tuning, and on-robot RL with expert corrections and rewards. PI reports more than 2x throughput on the hardest tasks and 2x or lower failure rates, with all-day espresso making, hours of laundry folding in a new home and factory box assembly.
| Architecture | π0.6 VLA (5B-parameter vision-language model augmented with an action expert, supporting heterogeneous prompts and conditioning) trained with Recap: a learned value function provides advantages, and the policy is advantage-conditioned. |
|---|
| Params | 5B-parameter VLM backbone + action expert (total not disclosed) |
|---|
| Inputs outputs | Inputs: images, text commands and conditioning metadata (execution quality, desired advantage); outputs: robot actions. |
|---|
| Training data | Offline-RL pretraining (Recap stage 1), task fine-tuning on demonstrations, then on-robot autonomous experience with expert teleoperated corrections and reward feedback. |
|---|
| Embodiments supported | PI robot platforms used for espresso making, laundry folding and box assembly. |
|---|
| License weights | Proprietary; not released. |
|---|
| Benchmarks claimed | | Benchmark | Throughput on hardest tasks (e.g. espresso) vs base model |
|---|
| Result | more than 2x |
|---|
| Source url | pi.website/…/pistar06 |
|---|
|
|---|
| Sources | |
|---|
Benchmarks, where listed, are as claimed by the vendor.
