π0.7: A Steerable Robotic Foundation Model with Emergent Compositional Generalization

TL;DR

π0.7 is Physical Intelligence’s next-generation generalist robot policy, positioned as a single model that matches RL-finetuned specialists (their prior π*0.6 “Recap” models) while showing the first signs of compositional generalization across tasks, robot embodiments, and even new appliances absent from training data. The recipe is diverse-context prompting: at training time the model ingests a mixture of language descriptions, sub-step language coaching, control-modality labels, “how-it-was-done” metadata, and visual subgoal images, which lets disparate data sources (multi-robot teleop, human videos, autonomous rollouts) be unified. At test time the same prompt slots accept either human-authored language coaching or visual subgoals produced by an auxiliary lightweight world model that hallucinates “what the next sub-step should look like.” Headline emergent behaviors: zero-/few-shot air-fryer operation from language coaching alone, and zero-shot laundry folding on a bimanual UR5e for which no UR5e folding data was collected.

Key claims

π0.7 follows new language commands and performs tasks never seen in training, showing initial signs of compositional generalization in a VLA — combining skills (loading objects, opening containers) to operate appliances like air fryers without dedicated demonstrations [§“What makes π0.7 generalize”].
The training recipe is diverse-context prompting: a single multimodal prompt structure accepts diverse language, “how-it-was-done” metadata (speed/quality), control-modality labels (joint vs end-effector), and visual subgoal images, which lets the model absorb heterogeneous data sources that would otherwise conflict [§“Diverse prompts unlock diverse data”].
Suboptimal autonomous-rollout data — which would ordinarily contaminate behavior cloning — can be safely incorporated by annotating it with lower-quality / lower-speed metadata, so the model learns to condition on quality rather than imitate it indiscriminately [§“Diverse prompts unlock diverse data”].
A lightweight auxiliary world model generates visual subgoal images at test time from language sub-tasks; these images are then fed back as part of π0.7’s multimodal prompt to drive low-level action prediction [§“Air fryer experiment”].
A separate fine-tuned high-level policy (extending the HiRobot framework) can autonomously emit the language sub-step sequence for a long-horizon task, closing the loop so π0.7 executes appliance use without human verbal coaching [§“Air fryer experiment”].
Cross-embodiment generalization: π0.7 folds laundry on a bimanual UR5e at a success rate matching the zero-shot success rate of expert human teleoperators (mean 375 h experience) attempting the task on the UR5e for the first time, despite no UR5e folding data in training [§“Cross-embodiment generalization”].
A single π0.7 model performs all of Recap’s previously specialist-trained tasks (laundry folding, espresso making, box folding) at similar or higher throughput than the per-task RL-trained specialists — by distilling autonomous Recap-rollout data into π0.7 with strategy metadata [§“Matching specialists”].
Knowledge of unseen appliances (e.g. how to operate an air fryer) appears to come from a combination of a small number of related robot episodes (DROID dataset, two home episodes of closing air fryers) and web-scale vision-language pre-training; the authors could not isolate a single source [§“Air fryer experiment”].

Method

The model is a vision-language-action transformer that ingests a multimodal prompt at every action-prediction step. The prompt slots are deliberately heterogeneous: (1) primary language describing the task; (2) optional fine-grained sub-step language (“coaching”) similar to how a human would guide a beginner; (3) metadata strings describing how the trajectory was performed (speed bucket, quality bucket); (4) control-modality tags (joint vs end-effector); (5) zero or more visual subgoal images showing what the end of the current sub-step should look like. This lets a single weight set absorb data from multiple robot platforms, human-video corpora, and autonomous rollouts of prior policies — including known-suboptimal autonomous data, which is admitted by labeling it as such rather than filtered out. At inference the same slots are populated by either a human user (zero-shot or with language coaching), by a fine-tuned high-level HiRobot-style policy that emits sub-step language autonomously, and/or by a lightweight visual-subgoal world model that produces a target image for each language sub-task. Recap-distilled strategy metadata is what lets π0.7 match per-task RL specialists from a single generalist model.

Results

One general-purpose π0.7 matches or exceeds task-specific Recap specialists on laundry folding, espresso making, and box folding (normalized throughput ≥ 1.0 against each specialist baseline; chart in §“Matching specialists” is reported qualitatively, no table of raw numbers).
Cross-embodiment zero-shot: bimanual UR5e t-shirt folding succeeds at a rate matching expert teleoperator zero-shot baselines (~375 h experience), without UR5e folding training data.
New-appliance qualitative results: zero-shot air-fryer loading fails to complete the task; with step-by-step language coaching the task completes effectively; with a fine-tuned high-level policy + visual subgoals, the loop runs fully autonomously.
No quantitative benchmark scores (e.g., on RoboCasa, LIBERO, CALVIN, BEHAVIOR) and no parameter counts, training compute, or training data size are disclosed in the blog post. The companion PDF is the place to look for numbers; the blog itself reports only normalized-against-specialist plots and qualitative video evidence.

Why it’s interesting

π0.7 lands in the wiki’s World Foundation Models cluster from a different direction than the other entries. Where NVIDIA bundles a WFM stack as a backbone under a VLA (NVIDIA Unveils New Open Models, Data and Tools to Advance AI Across Every Industry — Cosmos Reason 2 under Isaac GR00T N1.6), and Waymo specializes a closed flagship into a vertical simulator (The Waymo World Model: A New Frontier For Autonomous Driving Simulation), π0.7 inverts the dependency: the VLA is primary and a lightweight subgoal world model is consumed by the policy at inference, generating visual targets for the next sub-step. This is a third architectural pattern — “world model as on-policy advice” rather than “world model as backbone” or “world model as product surface.” It also sharpens The flavor of the bitter lesson for computer vision‘s open question about how action-conditional WFMs get trained without paired data: π0.7 sidesteps the chicken-and-egg by training the policy and the subgoal generator separately, with the generator conditioning on language sub-tasks rather than actions. Worth tracking against NVIDIA Unveils New Open Models, Data and Tools to Advance AI Across Every Industry as the open question of whether VLA progress will come from bigger WFM backbones or from richer prompt structures over the same backbone.