A side-by-side comparison of Physical Intelligence's VLA releases through model, data, and training lenses. Each version is an additive delta on the previous one — the table makes the deltas legible.

flowchart TB
  pi0["π0 (2024)<br/>Cross-embodiment<br/>flow-matching VLA"]
  pi05["π0.5 (2025)<br/>+ co-training<br/>+ subtask hierarchy<br/>+ FAST-token pretrain<br/>→ open-world generalization"]
  pi06["π0.6 (Nov 2025)<br/>+ Gemma3-4B + KI<br/>+ optional metadata<br/>→ out-of-the-box dexterity"]
  pistar["π*0.6 (Nov 2025)<br/>+ RECAP<br/>(advantage-conditioned RL)<br/>→ specialist-grade policies"]
  pi07["π0.7 (Apr 2026)<br/>+ MEM history + subgoal<br/>  images + metadata<br/>+ suboptimal + RL data<br/>→ compositional generalization"]
  pi0 --> pi05 --> pi06
  pi06 --> pistar
  pi06 --> pi07
  pistar -. rollouts distilled .-> pi07

• π0 → π0.5: Hierarchical split (high-level subtask head + low-level action expert). Discrete FAST-token stage for pretraining.
• π0.5 → π0.6: Backbone upgraded to Gemma3-4B, action expert grown to 860M, Knowledge Insulation stabilizes joint training.
• π0.6 → π0.7: Keeps the same backbone/expert; adds MEM video-history encoder, subgoal-image conditioning, linear-projected proprioception, RTC, and CFG on metadata. Worldmodel (BAGEL-14B) is a separate module that feeds the prompt, not an architectural change to the policy itself.

• π0 → π0.5: Large-scale home data (~400 hrs MM × 100 homes), web multimodal co-training.
• π0.5 → π0.6: More of the same, richer metadata annotation.
• π0.6 → π0.7: Doubles down on heterogeneity — failures, autonomous rollouts, RL-specialist traces, egocentric human video, DROID. The novelty is not the sources but that metadata conditioning lets them be used productively instead of averaged.

• π0 → π0.5: Two-stage FAST→flow-matching; co-training on VQA/detection/subtask text.
• π0.5 → π0.6: Knowledge Insulation (one-stage joint training with gradient-insulated VLM).
• π0.6 → π*0.6: Advantage-conditioned RL (RECAP) — no log-prob RL needed (flow-matching policies give no action log-probs, so PPO/SAC are inapplicable). A multi-task distributional value function $V^{\pi_{ref}}(o_t,\ell)$ is trained on all data (demos + autonomous rollouts + teleoperated interventions); the policy then conditions on a binarized advantage indicator as a text token ("Advantage: positive/negative"), and at inference is run on the positive branch. Trained end-to-end with Knowledge Insulation. On the hardest tasks (laundry folding, box assembly, espresso) RECAP more than doubles task throughput and roughly halves the failure rate vs the base policy.
• π0.6 → π0.7: Rich prompt conditioning + per-component dropout + CFG, distilling π*0.6 RL behavior into a generalist without RL at post-train time.

The π series is evolving along a consistent thesis: push more capability into pre-training so post-training becomes optional. π0 needed per-task finetuning; π0.5 reduced it; π0.6 removed it for many tasks; π0.7 removes it even for tasks previously only solvable by RL specialists — and shows the first signs of compositional (not just in-distribution) generalization in the series (the π0.7 blog phrases it as "the first signs of compositional generalization," recombining skills across tasks).

• π0.6
• π0.7 🆕
• π*0.6 + RECAP
• Survey: VLA & Manipulation

← Back to Home