The Pi Line

The VLA module ended on a bottleneck: discrete action tokens let a web-pretrained model emit actions with the machinery it already has, but autoregressive decoding caps the control rate an order of magnitude below what dexterous manipulation needs. Physical Intelligence's pi line answers it with a different split of labor. Keep the pretrained backbone, replace the token head with a flow-matching action expert, and then keep fixing what that first design got wrong. This module walks the line from π0 (October 2024) to π0.7 (April 2026), including the point where the open weights stop.

π0: the action expert

π0 grafts a second transformer onto a pretrained vision-language model Black 2024π0: A Vision-Language-Action Flow Model for General Robot ControlKevin Black, Noah Brown, Danny Driess et al., RSS 2025, 2024. The first set of weights handles image and text tokens and is initialized from PaliGemma (3B). The second, the action expert, is initialized from scratch and adds 300M parameters, for 3.3B total. The structural point matters for anyone implementing one: the expert is a Mixture-of-Transformers, not an MLP head. It is a full transformer whose 50 action tokens attend into the backbone's activations at every layer through shared attention, with a bidirectional mask among themselves. Observations are 2 or 3 RGB images, tokenized language, and proprioceptive state; the output is a chunk of H = 50 actions, one second at the full 50 Hz control rate Black 2024π0: A Vision-Language-Action Flow Model for General Robot ControlKevin Black, Noah Brown, Danny Driess et al., RSS 2025, 2024.

Where Diffusion Policy denoises with a DDPM, π0 trains the expert with conditional flow matching on a straight transport path. Sample noise $\varepsilon \sim \mathcal{N}(0, I)$, form the interpolant $A_t^\tau = \tau A_t + (1 - \tau)\varepsilon$, and regress the network $v_\theta$ onto the target vector field $u = A_t - \varepsilon$, with the training timestep $\tau$ drawn from a beta distribution weighted toward the noisy end. At inference, integrate from $\tau = 0$ to $\tau = 1$ with forward Euler in 10 steps; the image-text-state prefix is KV-cached once, so each step recomputes only the action-token suffix, which is what makes 10 steps affordable Black 2024π0: A Vision-Language-Action Flow Model for General Robot ControlKevin Black, Noah Brown, Danny Driess et al., RSS 2025, 2024.

Cross-embodiment is handled crudely and effectively: one padded action and state vector sized for the widest robot, with per-embodiment normalization statistics, so a single checkpoint serves single-arm, bimanual, and mobile platforms Black 2024π0: A Vision-Language-Action Flow Model for General Robot ControlKevin Black, Noah Brown, Danny Driess et al., RSS 2025, 2024. Training splits into broad pre-training and task-specific post-training, mirroring LLM practice. The limitations set the agenda for everything after: execution was synchronous (finish the chunk, pause for inference, start the next), language following was mediocre, and hard tasks like laundry folding were near-zero success without fine-tuning Black 2024π0: A Vision-Language-Action Flow Model for General Robot ControlKevin Black, Noah Brown, Danny Driess et al., RSS 2025, 2024.

Why ten steps is enough

Flow matching's selling point over diffusion is path geometry. A DDPM trajectory curves through noise space; the rectified flow path from noise to action is nearly a straight line, so a coarse Euler integration tracks it well. The interactive below makes the trade concrete. Samples start as Gaussian noise in a 2D action space and are transported toward two action modes; the slider sets the integration budget from 1 to 50 steps.

At 1 or 2 steps the transported cloud visibly misses the modes: each Euler step samples the field once and cuts the corner on everything the field does between samples. By 5 to 10 steps the cloud concentrates on the targets, and pushing to 50 buys nothing measurable. That shape of curve is the whole deployment argument. Every step costs a forward pass through the expert, and at 50 Hz the budget is fixed by physics, not by GPU spend: π0 shipped 10 steps, π0.6 and π0.7 run 5, with π0.6 completing a full chunk inference in 63 ms on one H100 with three cameras Physical Intelligence 2025π0.6 Model CardPhysical Intelligence, 2025. The consistency and one-step distillation line covered in Diffusion Policy exists for the same reason.

π0-FAST: tokens done right

The discrete-token idea from RT-1 and OpenVLA turns out to fail for a specific, diagnosable reason: at 50 Hz, consecutive actions are nearly identical, so next-token prediction over naively binned actions becomes trivially easy and learns nothing useful. The FAST paper reports that standard discretization "fails completely" on high-frequency dexterous data Pertsch 2025FAST: Efficient Action Tokenization for Vision-Language-Action ModelsKarl Pertsch, Kyle Stachowicz, Brian Ichter et al., RSS 2025, 2025.

FAST tokenizes the action chunk in the frequency domain instead. Apply a discrete cosine transform along the time axis of each action dimension, quantize the coefficients (most of the energy sits in the low frequencies, so high-frequency coefficients quantize to zero), then byte-pair-encode the sparse integer sequence. The result is a short, dense token stream an autoregressive VLM predicts with ordinary cross-entropy. The released FAST+ variant is a universal tokenizer trained on 1M real robot trajectories, usable as a black box across action spaces Pertsch 2025FAST: Efficient Action Tokenization for Vision-Language-Action ModelsKarl Pertsch, Kyle Stachowicz, Brian Ichter et al., RSS 2025, 2025.

π0-FAST, the FAST tokenizer paired with the π0 backbone, matches the flow-matching original on 10k hours of robot data while cutting training time by up to 5x Pertsch 2025FAST: Efficient Action Tokenization for Vision-Language-Action ModelsKarl Pertsch, Kyle Stachowicz, Brian Ichter et al., RSS 2025, 2025. The catch is inference: autoregressive decoding of 30 to 60 tokens made π0-FAST roughly 2x slower to complete tasks than the flow-matching π0 in a head-to-head real-robot comparison Physical Intelligence 2025Knowledge InsulationPhysical Intelligence, 2025. FAST's lasting contribution turned out to be not the deployed decoder but the training objective, which is where π0.5 picks it up.

π0.5 and Knowledge Insulation

π0.5 targeted generalization rather than speed Black 2025π0.5: a Vision-Language-Action Model with Open-World GeneralizationKevin Black, Noah Brown, James Darpinian et al., CoRL 2025, 2025. The architecture keeps the π0 skeleton, but the training mixture adds hybrid multimodal examples that interleave image observations, language commands, object detections, semantic subtask predictions, and low-level actions in a single sequence. The subtask prediction makes the model internally hierarchical: at inference it first predicts a high-level language subtask at low frequency ("pick up the plate"), then conditions the action expert on that subtask at high frequency. There is no separate planner. The payoff was the first end-to-end system shown cleaning kitchens and bedrooms it had never seen, and the paper's ablation shows the non-action data is essential to that result, not incremental Black 2025π0.5: a Vision-Language-Action Model with Open-World GeneralizationKevin Black, Noah Brown, James Darpinian et al., CoRL 2025, 2025.

The load-bearing training trick arrived shortly after, under an unglamorous name. Knowledge Insulation starts from a diagnosis: gradients from a from-scratch, continuous-generative action expert corrupt the pretrained backbone's language processing. The paper's example is a model told to put a spoon in the dish container that grabs the trash instead. The fix has three parts Driess 2025Knowledge Insulating Vision-Language-Action Models: Train Fast, Run Fast, Generalize BetterDanny Driess, Jost Tobias Springenberg, Brian Ichter et al., NeurIPS 2025, 2025:

1. Train the action expert with flow matching on the continuous actions that actually execute.
2. Train the backbone with discrete FAST action tokens, cross-entropy on a representation the VLM's pretraining objective already likes, plus web vision-language data and subtask data.
3. Stop the expert's gradients at the interface. The expert still attends into backbone activations in the forward pass; only the backward pass is severed.

The ablation that makes the recipe non-obvious: stop-gradient alone is not enough. A fully insulated backbone that never sees robot data lacks representations adapted for motor control, and the expert cannot use them (the shirt-folding task is unsolvable in that configuration). The FAST-token supervision adapts the backbone for control without exposing it to the denoising loss that damaged it Driess 2025Knowledge Insulating Vision-Language-Action Models: Train Fast, Run Fast, Generalize BetterDanny Driess, Jost Tobias Springenberg, Brian Ichter et al., NeurIPS 2025, 2025. With the full recipe, π0.5 reaches a given performance level on the bussing task in 7.5x fewer training steps than π0, keeps flow matching's fast inference, and beats both π0 and π0-FAST on language following Driess 2025Knowledge Insulating Vision-Language-Action Models: Train Fast, Run Fast, Generalize BetterDanny Driess, Jost Tobias Springenberg, Brian Ichter et al., NeurIPS 2025, 2025.

π0.6 and π*0.6: the line goes closed

π0.6, released November 2025 as a model card rather than an arXiv paper, scales the recipe up Physical Intelligence 2025π0.6 Model CardPhysical Intelligence, 2025. The backbone initializes from Gemma3 4B behind a SigLIP 400M vision encoder; the action expert grows to 860M parameters with the same layer count as the backbone; the model consumes up to four 448x448 images per step. Training uses Knowledge Insulation end to end, and inference runs 5 denoising steps. The headline result is reliability without task-specific post-training: out of the box, π0.6 folds laundry reliably and fully assembles a box 20% of the time, both near-zero for π0.5 without fine-tuning Physical Intelligence 2025π0.6 Model CardPhysical Intelligence, 2025.

π*0.6 layers reinforcement learning on top via the Recap recipe, which mirrors how a person learns a manual skill: demonstrations define the behavior, coaching (an expert teleoperator taking over mid-rollout) corrects the mistakes the policy actually makes, and practice (autonomous rollouts) compounds Physical Intelligence 2025π*0.6: a VLA that Learns from ExperiencePhysical Intelligence, 2025. The mechanism that makes RL tractable on a 5B flow-matching VLA is advantage conditioning. A language-conditioned value function predicts steps-to-completion, solving credit assignment (if a portafilter grasped at a bad angle fails the insertion 20 seconds later, the value function attributes the blame to the grasp); the advantage is binarized and fed back to the VLA as a conditioning token, so all data, good and bad, stays in training and the model is simply told which is which. At execution you condition on "high advantage." Espresso throughput and success more than doubled, with greater than 90% success on all three demonstrated applications Physical Intelligence 2025π*0.6: a VLA that Learns from ExperiencePhysical Intelligence, 2025.

A March 2026 variant, π0.6-MEM, added two-scale memory: a video encoder with interleaved spatial and causal-temporal attention for short-term observation history, and model-authored natural-language notes for long-term state. That combination supports tasks requiring up to 15 minutes of context and produces in-context adaptation, like retrying a failed grasp differently instead of identically Torne 2026VLAs with Long and Short-Term MemoryMarcel Torne, Karl Pertsch, Homer Walke et al., 2026.

All three are closed: no weights and no arXiv papers, only the model cards and lab PDFs linked here as primary sources.

π0.7: generalization as a prompting problem

π0.7 (April 2026) reframes the generalization problem: the bottleneck is underspecified prompts, not model capacity Physical Intelligence 2026π0.7: a Steerable Model with Emergent CapabilitiesPhysical Intelligence, 2026. The skeleton is π0.6 plus MEM at about 5B parameters, with two architectural tweaks worth noting: the flow timestep is injected through adaptive RMSNorm, and proprioceptive state enters through a linear projection rather than discretized text tokens. Training-time Real-Time Chunking with simulated delays of 0 to 12 timesteps lets the policy tolerate up to 240 ms of inference latency on a 50 Hz robot Physical Intelligence 2026π0.7: a Steerable Model with Emergent CapabilitiesPhysical Intelligence, 2026.

The actual contribution is the context every training example carries, with each component randomly dropped so any subset works at test time Physical Intelligence 2026π0.7: a Steerable Model with Emergent CapabilitiesPhysical Intelligence, 2026:

• Task and subtask language, the latter produced at runtime by a learned high-level policy or a human coaching live.
• Episode metadata: quality score, speed, mistake flags, desired length. Labeling suboptimal autonomous data as low quality makes it safe to train on, because the model learns it as the "low quality" mode rather than as the target behavior. At test time the metadata is set to ideal values.
• A control-modality label (joint versus end-effector control).
• Multi-view subgoal images, generated at runtime by a lightweight world model initialized from BAGEL, refreshed asynchronously so VLA inference never blocks on it.

The reported results are the strongest generalization claims in the line: zero-shot laundry-folding transfer to a bimanual UR5e (an embodiment with no laundry data) matching first-time expert human teleoperators, and compositional tasks like loading an air fryer, which the model attempts zero-shot, completes with step-by-step language coaching, and then learns autonomously from the coaching transcripts with no additional teleoperation Physical Intelligence 2026π0.7: a Steerable Model with Emergent CapabilitiesPhysical Intelligence, 2026. Read the compositional claim with the authors' own hedge: they describe "initial signs," and the zero-shot air fryer attempt fails without coaching. Weights are closed.

Where the open weights stop

The openpi repository hosts π0, π0-FAST, and π0.5 Physical Intelligence 2024openpiPhysical Intelligence, 2024. Everything after, π0.6, π*0.6, π0.6-MEM, and π0.7, is closed, documented only by lab PDFs and blog posts. The timeline below is the line at a glance; select any generation for its backbone, contribution, and primary source.

The arc of the line is a useful compression of the whole domain's recent history. π0 established the action expert as the standard unit. π0-FAST fixed tokenization and accidentally produced the objective that made the backbone trainable. π0.5 bought open-world generalization with heterogeneous co-training, and Knowledge Insulation made the recipe efficient. The closed generations then spent their budget on what the open line still lacks: reliability without fine-tuning, learning from experience, memory, and prompts rich enough to absorb suboptimal data. The next module steps outside one lab and surveys the other generalist policies, open and closed, competing with this line.