VLA GalleryAll modelsGallery

VLA Gallery / All models / π family

π0.5 + KI

π0.5 + KI (Knowledge Insulation)

Physical Intelligence · 2025-05

VLAπ familyTotal not disclosedhybrid

Blocks action-expert gradients from flowing into the VLM to preserve VLM knowledge while learning representations with FAST tokens.

View architecture diagram in the galleryPaper ↗

Paper

Knowledge Insulating Vision-Language-Action Models: Train Fast, Run Fast, Generalize BetterarXiv 2505.23705

Overview

Category
π family
Type
VLA
Released
2025-05
Organization
Physical Intelligence
Open weights
Partial
Parameters
Total not disclosed; 2B LM backbone (width 2048, depth 18, mlp 16,384) + 300M action expert

Architecture

Backbone LLM
Gemma 2B · PaliGemma config 2B VLM
Vision encoder
SigLIP 400M
Action head
FAST tokens (VLM) + flow expert 300M, stop-gradient (hybrid)
Chunk
50 steps
Control rate
Not disclosed
Architecture (original summary)
PaliGemma-initialized 2B LM backbone + 300M action expert; VLM trained on FAST discrete tokens, expert on flow matching (horizon 50), with stop-gradient blocking expert gradients from reaching the VLM; co-trained on VLM data

Training

Pretraining accelerators
Not disclosed
Pretraining duration
Not disclosed
Pretraining steps & batch
Not disclosed (π0 needs 7.5× the steps to reach similar performance)
GPU/TPU-hours
Not disclosed
Fine-tuning
Not disclosed
Data type
Real-robot teleoperation + web vision-language
Data scale
Not disclosed (12 embodiment configs + OXE + VLM data)
Main datasets
PI's own data, OXE; CapsFusion, COCO, Cambrian-7M, PixMo, VQAv2, object localization data
Tasks & evaluation
LIBERO (from generalist 98.0/97.8/95.6/85.8, LIBERO-90 96.0; from scratch LIBERO-90 92.7); DROID 0.55±0.09 (π0 0.49±0.09, π0-FAST 0.45±0.09); drawer insertion·bussing·shirt folding, 4 mobile tasks in unseen homes (10 episodes per task)
Robot
12 configurations: single-arm ARX·UR5·Franka; bimanual ARX·AgileX·Trossen·UR5; bimanual mobile Trossen·ARX slate·Galaxea G1·Hexmove H1·Fibocom

Inference

Inference device & speed
Own latency figures not disclosed; cites π0-FAST ≈750 ms (RTX 4090), π0-style 10 Hz vs autoregressive VLA 1.3 Hz
Inference memory
Not disclosed
Weights bf16 (calc.)
Not disclosed
Where it runs
Not disclosed

Reinforcement learning & base

Reinforcement learning
Not used · Supervised learning (VLM protected via stop-gradient)
Base control
Wheeled (Format not stated) · Mobile bimanual tasks (Make bed·Laundry in basket, etc.) include movement; base action dimension layout Not stated in paper (π0.5 approach)

Notes

Key features
Blocks action-expert gradients from flowing into the VLM to preserve VLM knowledge while learning representations with FAST tokens. π0 needed 7.5× more steps to reach similar performance, and this method adds about 20% compute per step.
Notes
'π0.5 + KI' is a task-assigned label; the paper calls it 'knowledge insulation'. openpi's π0.5 and pi05_droid are trained with KI. Appendix B num_heads=18 conflicts with 8 in openpi. π0-FAST takes 2× longer execution time on the drawer task.
Verification
Confirmed

Other models in this category