Venue: Physical Intelligence (technical report) · Date: April 16, 2026 Category: Baseline (VLA Architecture + Data) Trend tag: Baseline for Trend 1 (supersedes π0.6)

flowchart LR
  subgraph Prompt[Multimodal Prompt Ct]
    L[Task instruction<br/>e.g. 'clean the kitchen']
    SL[Subtask ℓ̂<br/>e.g. 'open fridge']
    SG[Subgoal images gt<br/>multi-view]
    M[Metadata<br/>speed / quality 1-5 / mistake]
    C[Control mode<br/>joint or EE]
  end
  V[4 cameras × up to<br/>6 history frames<br/>448×448] --> MEM[MEM-style video<br/>history encoder]
  MEM --> B[Gemma3-4B<br/>VLM backbone]
  Prompt --> B
  PR[Proprioception qt<br/>linear projection] --> B
  B --> AE[860M flow-matching<br/>Action Expert<br/>50-step chunk]
  noise[Noise] --> AE
  AE -- 5 Euler steps --> A[Continuous action chunk]
  A -- RTC, 240ms tolerance --> R[Real robot]

  HL[High-level policy<br/>same arch] -.generates.-> SL
  WM[World model gψ<br/>BAGEL-14B MoT] -.generates.-> SG

Prior VLAs (π0, π0.5, π0.6) were trained on high-quality demos with minimal prompt context — just a task string. This produces generalists that need task-specific RL/SFT post-training (π*0.6) to match specialist performance, and that fail to compositionally generalize to unseen combinations (new appliances, new embodiments). Naively mixing failed/suboptimal/autonomous data into training just averages modes and degrades performance.

Core idea: expand the training context Ct from a plain instruction to a rich multimodal prompt so the model can absorb heterogeneous data without mode collapse:

• Subtask text ℓ̂ — intermediate semantic step (inherited from π0.5).
• Subgoal images — multi-view near-future images, generated at runtime by a BAGEL-14B world model that is web-pretrained on image editing/video.
• Episode metadata — discretized speed, quality score 1–5, mistake flag. At test time, always prompt "quality = 5, mistake = false" to steer toward high-quality modes.
• Control mode — joint vs. end-effector label. Each component is randomly dropped out at train time (subgoal kept in 25%, subtask in 70% of those, metadata dropped 15%), so any subset can be used at inference. Classifier-free guidance on metadata further amplifies high-quality behavior.

Architecture: ~5B total = Gemma3-4B VLM backbone + MEM video-history encoder (temporal+spatial compression, fixed token count regardless of #frames) + 860M flow-matching action expert (50-token chunk, 5 Euler steps). Inherits Knowledge Insulation from π0.6 (VLM trained on FAST tokens + web; action-expert gradients don't propagate back). Training-time Real-Time Chunking (RTC) simulates 0–240ms inference delay for smooth deployment.

Data: the prompt-conditioning is what unlocks the real novelty — heavy use of suboptimal data: lower-quality demos, failure episodes, autonomous rollouts from π*0.6 RL training, human egocentric video, open-source robot data (DROID etc.), and web multimodal (VQA, detection, captioning). Metadata labels disambiguate each episode's quality so the model can distill RL-specialist behavior rather than average it away.

On the four π*0.6 RL-specialist tasks (laundry, diverse laundry, box building, espresso), a single general-purpose π0.7 matches or exceeds specialist success rate and throughput — no task-specific post-training. Also matches SFT specialists on shirt-inside-out, PB sandwich, drive-through-door, slice zucchini, peel, trash. Zero-shot cross-embodiment: π0.7 folds t-shirts on a UR5e (never trained for laundry on UR5e) at the level of expert teleoperators — 85.6% task progress / 80% success vs. 10 experienced teleoperators' 90.9% / 80.6% — discovering embodiment-adapted strategies (e.g. single-arm or vertical grasps) rather than copying source behaviors. Compositional generalization: operates a never-seen air fryer (loading a sweet potato, unloading, toasting a bagel) with zero air-fryer training episodes in the robot data — only similar appliances seen in human and external datasets — by being coached through with step-by-step language. Ablations: removing metadata or eval-data both hurt — diverse data and detailed context are synergistic.

Shifts the frontier of the π series from "post-train per task" to "one prompt-steerable generalist." Three structural moves:

1. Prompt as the integration surface — instead of changing architecture for new capabilities, package them as extra prompt modalities (à la LLM prompt expansion).
2. Suboptimal-data-as-a-feature — RL rollouts and failures become training signal via metadata conditioning, sidestepping the π*0.6 + RECAP detour for many tasks.
3. World-model subgoal images bring visual grounding into the prompt, complementing subtask text when language is ambiguous. Positions π0.7 as the new production baseline against which ICLR 2026's discrete-diffusion and memory VLAs will be compared, and gives the first π-series paper with strong claims of compositional (not just in-distribution) generalization.

• Blog / project: https://www.pi.website/blog/pi07
• Paper PDF: https://www.pi.website/download/pi07.pdf
• arXiv (v1 Apr 16 2026, v2 Apr 24 2026): https://arxiv.org/abs/2604.15483
• Project page: https://pi.website/pi07
• π0.6 model card: https://website.pi-asset.com/pi06star/PI06_model_card.pdf
• π*0.6 (RECAP) arXiv: https://arxiv.org/abs/2511.14759
• π0.5 arXiv: https://arxiv.org/abs/2504.16054
• π0 arXiv: https://arxiv.org/abs/2410.24164
• OpenPi: https://github.com/Physical-Intelligence/openpi
• TechCrunch coverage: https://techcrunch.com/2026/04/16/physical-intelligence-a-hot-robotics-startup-says-its-new-robot-brain-can-figure-out-tasks-it-was-never-taught/

For a long-form review with full results tables, all ablations, and limitations: In-Depth Review of π0.7.

• π series evolution — side-by-side model/data/training comparison across π0 → π0.7
• π0.6 (direct predecessor)
• π*0.6 + RECAP (the RL sibling whose specialists π0.7 now distills)
• MemoryVLA (memory axis — π0.7 uses MEM-style history encoder)
• Discrete Diffusion VLA (the architectural alternative)
• Survey: VLA & Manipulation

← Back to ICLR-2026