π0.7

Physical Intelligence · 2026-04

VLANew in 2026About 5Bflow matching

Its key feature is diverse context conditioning that adds execution quality·speed·strategy metadata and subgoal images to language.

Paper

Category: New in 2026
Type: VLA
Released: 2026-04
Organization: Physical Intelligence
Open weights: No
Parameters: About 5B (Gemma3 4B + action expert 860M; subgoal world model BAGEL 14B separate)

Backbone LLM: Gemma 3 4B
Vision encoder: SigLIP (Gemma 3)
Action head: flow expert 860M (5 steps) (flow matching)
Chunk: 50 (15 · 25 executed)
Control rate: Not disclosed
Architecture (original summary): Gemma3 4B VLM + 860M flow matching action expert + MEM-style video history encoder; 50-step chunks, 5 denoising steps, execute 15·25 steps; joint and end-effector control

Pretraining accelerators: Not disclosed (no GPU/TPU statement in paper)
Pretraining duration: Not disclosed
Pretraining steps & batch: Not disclosed
GPU/TPU-hours: Not disclosed
Fine-tuning: Not disclosed
Data type: Real-robot teleoperation (multiple platforms) + autonomous collection (failure episodes, π*0.6 RL data) + human video (egocentric) + web vision-language + open-source robot data
Data scale: Not disclosed (hours·episodes·mixture ratio)
Main datasets: In-house multi-embodiment demonstrations, π*0.6 RL data, open-source robot datasets (incl. DROID, secondary)
Tasks & evaluation: Espresso machine, laundry folding, box assembly, trash bag removal, vegetable peeling, air fryer; bimanual UR5e T-shirt folding 80% without folding data (secondary); espresso on par with an RL-fine-tuned specialist model
Robot: Multiple platforms (fixed·mobile, single-arm·bimanual), bimanual UR5e

Inference device & speed: Simulates 0–12 timesteps of latency during training (up to 240 ms on a 50 Hz robot); inference hardware not disclosed
Inference memory: Not disclosed
Weights bf16 (calc.): 10 GB
Where it runs: Not disclosed

Reinforcement learning: Partial (RL-generated data) · No RL stage of its own. Distills using data collected by π*0.6 during RL training as extra examples
RL evidence: https://arxiv.org/html/2604.15483v2
Base control: Wheeled (Format not stated) · Includes mobile bimanual (3 DoF omnidirectional base) evaluation ('driving through a door'); actions applied via PD controller, base action format Not stated

Key features: Its key feature is diverse context conditioning that adds execution quality·speed·strategy metadata and subgoal images to language. Subgoal images are produced by a world model initialized from BAGEL 14B.
Notes: arXiv v1 2026-04-16, v2 2026-04-24. π blog could not be read (HTTP 403). No mention of weight release. Only 2 training episodes for closing the air fryer (secondary).
Verification: Confirmed