The Pi Line

The VLA module ended on a bottleneck: discrete action tokensaction tokenizationExpressing continuous robot actions as discrete tokens so a language model's machinery can produce them. RT-2 discretized each action dimension into 256 uniform bins and mapped the bin indices onto rarely used tokens of the model's existing vocabulary, which let the policy train with ordinary next-token prediction. The representation is simple and inherits the backbone's web knowledge, but 256 bins are coarse and autoregressive decoding is slow, the two weaknesses later work attacked with continuous experts and parallel decoding. 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-matchingflow matchingA generative modeling recipe that learns a time-dependent vector field carrying samples from a noise distribution to the data distribution, trained by regressing the model's field against the conditional flow rather than by estimating a score. pi0 brought it to robot control: a dedicated action expert is trained with the flow-matching objective to denoise continuous action chunks, and at inference the learned field is integrated as an ODE for a handful of steps, which keeps the policy fast enough for high-frequency control while keeping the multimodality that made diffusion policies attractive. 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 what the pinned checkpoint catalogue actually lists.

The whole design rests on one empirical curve. Drag the integration steps slider below from 1 upward: at one or two steps the transported cloud misses the action modes badly, by five to ten it has concentrated on them, and pushing to 50 buys nothing measurable. That is why a flow-matching head can close a 50 Hz loop where an autoregressive token head cannot.

π0: the action expert

π0 grafts a second transformer onto a pretrained vision-language modelvision-language modelA model pretrained jointly on web-scale image and text data, so visual recognition and language semantics live in one set of weights. RT-2's bet was that this pretraining is an asset for robot control: co-fine-tuning a vision-language model on robot trajectories and its original web data together transfers semantic knowledge, recognizing objects and following instructions the robot data never covered, into the policy. The vision-language-action models that followed all start from such a backbone. Black 2024↓π0: 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. The second set, the action expert, handles robotics-specific tokens and has approximately 300M parameters in the paper’s implementation. The structural point matters for anyone implementing one: the action expert is not an MLP head but a second set of transformer weights inside a single model, in the paper's words "a single transformer with two sets of weights (also known as experts)." It is a full transformer whose 50 action tokens attend into the backbone's activations at every layer through shared self-attention, with a bidirectional mask among themselves. Observations are 2 or 3 RGB images, tokenized language, and proprioceptive state; the model predicts H = 50 actions. The paper’s UR5e and Franka setups run at 20 Hz and execute 16 actions before another inference; other tested robots run at 50 Hz and execute 25 Black 2024↓π0: A Vision-Language-Action Flow Model for General Robot ControlKevin Black, Noah Brown, Danny Driess et al., RSS 2025, 2024. Whether the system then runs all fifty of those actions is a separate question: this is the predicted horizon, and how much of it a π release actually commits before the next inference is what the later sections of this module track.

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 2024↓π0: A Vision-Language-Action Flow Model for General Robot ControlKevin Black, Noah Brown, Danny Driess et al., RSS 2025, 2024.

Cross-embodimentcross-embodimentLearning from, or transferring to, robots with different morphologies: different joint counts, action dimensions, and sensor suites. The Open X-Embodiment project pooled data from 22 robot embodiments into a single training mixture and showed that the resulting RT-X policies carry skills across platforms. The gain is not free, because each body defines its actions differently, and the field answers that mismatch with several incompatible conventions. training combines data from robot types with different configuration spaces and action representations, including single-arm, dual-arm and mobile manipulators Black 2024↓π0: A Vision-Language-Action Flow Model for General Robot ControlKevin Black, Noah Brown, Danny Driess et al., RSS 2025, 2024. The paper separates broad pre-training from task-specific post-training: the base model performs tasks at what the authors call “rudimentary proficiency,” while complex tasks such as laundry folding use curated post-training data Black 2024↓π0: A Vision-Language-Action Flow Model for General Robot ControlKevin Black, Noah Brown, Danny Driess et al., RSS 2025, 2024. Its v4 execution description commits 16 of the 50 predicted actions on UR5e/Franka and 25 on the other tested robots, rather than a universal 50-step commitment. The RTC paper later evaluates asynchronous execution against stopping between chunks using π0.5; that experiment is not evidence that every original π0 deployment executed all 50 predictions Black 2025↓Real-Time Execution of Action Chunking Flow PoliciesKevin Black, Manuel Y. Galliker, Sergey Levine, 2025.

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. That is what the figure at the top of this module measures: samples start as Gaussian noise in a 2D action space and are transported toward two action modes under a fixed integration budget.

Each Euler step samples the field once and cuts the corner on everything the field does between samples, which is why one or two steps miss. The shape of that 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 uses 10 integration steps Black 2024↓π0: A Vision-Language-Action Flow Model for General Robot ControlKevin Black, Noah Brown, Danny Driess et al., RSS 2025, 2024. π0.6 reports five steps and 63 ms per chunk on one H100 with three cameras Physical Intelligence 2025↓π0.6 Model CardPhysical Intelligence, 2025. π0.7 also uses five denoising steps in its experiments Ai 2026↓π0.7: a Steerable Generalist Robotic Foundation Model with Emergent CapabilitiesBo Ai, Ali Amin, Raichelle Aniceto et al., 2026. Those step counts and π0.6’s hardware-specific chunk latency do not measure a common robot-control frequency. The consistency and one-step distillation line covered in Diffusion Policy exists for the same reason.

π0-FAST: tokens done right

Pertsch 2025↓FAST: Efficient Action Tokenization for Vision-Language-Action ModelsKarl Pertsch, Kyle Stachowicz, Brian Ichter et al., 2025 FAST diagnoses a weakness of simple per-dimension, per-timestep binning: highly correlated action tokens can make next-token prediction too easy, for example by copying a previous token. In the paper’s evaluations, policies trained with this naïve tokenization make no progress on table bussing at 20 Hz or T-shirt folding at 50 Hz. This is a result for those training setups, not a claim that every discretization fails.

FAST normalizes each action dimension, applies a discrete cosine transform along time, scales and rounds the coefficients, and byte-pair-encodes their flattened sequence. Rounding trades reconstruction fidelity for compression; the BPE stage losslessly compresses the quantized integers. The universal FAST+ tokenizer is trained on approximately one million one-second action chunks, spanning different embodiments, action spaces and control frequencies. The abstract calls these trajectories, but Section V-C states the more precise chunk unit Pertsch 2025↓FAST: Efficient Action Tokenization for Vision-Language-Action ModelsKarl Pertsch, Kyle Stachowicz, Brian Ichter et al., 2025.

The paper trains π0-FAST on a 10k-hour cross-embodied robot-data mixture. In its generalist comparison, π0-FAST matches the compared π0 task performance with five times fewer GPU hours of training. The abstract summarizes the benefit as "up to 5x"; it is a training-compute result, not an inference speedup or a universal training-step ratio Pertsch 2025↓FAST: Efficient Action Tokenization for Vision-Language-Action ModelsKarl Pertsch, Kyle Stachowicz, Brian Ichter et al., 2025. Inference is a separate trade-off. The FAST paper reports approximately 750 ms per one-second action chunk on an NVIDIA 4090 for π0-FAST, versus within 100 ms for its diffusion π0 comparison; decoding typically requires 30 to 60 action tokens instead of ten diffusion steps. Separately, the Knowledge Insulation note reports that π0-FAST took twice as long to solve the single-embodiment table-bussing task in its comparison. That is task wall-clock time, not a generic twofold inference ratio Pertsch 2025↓FAST: Efficient Action Tokenization for Vision-Language-Action ModelsKarl Pertsch, Kyle Stachowicz, Brian Ichter et al., 2025 Driess 2025↓VLAs that Train Fast, Run Fast, and Generalize BetterDanny Driess, Jost Tobias Springenberg, Brian Ichter et al., 2025. The later KI recipe uses FAST tokens for backbone training while retaining continuous-action inference.

π0.5 and Knowledge Insulation

π0.5 targeted generalization rather than speed Black 2025↓π0.5: a Vision-Language-Action Model with Open-World GeneralizationPhysical Intelligence, Kevin Black, Noah Brown et al., 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 paper evaluates long-horizon mobile manipulation in new home environments, and the paper's ablations show the heterogeneous mixture carries the result: removing either cross-embodiment data source significantly degrades the home-task scores, while web data, whose removal was not statistically significant there, has a large effect on language following and high-level subtask inference Black 2025↓π0.5: a Vision-Language-Action Model with Open-World GeneralizationPhysical Intelligence, Kevin Black, Noah Brown et al., 2025.

Driess 2025↓Knowledge Insulating Vision-Language-Action Models: Train Fast, Run Fast, Generalize BetterDanny Driess, Jost Tobias Springenberg, Brian Ichter et al., 2025 The Knowledge Insulation paper studies how a newly initialized continuous-action expert can interfere with pretrained VLM representations. Its Figure 2 shows π0 being instructed to bus a spoon into the bin but grasping trash instead. The authors extend the earlier two-stage π0.5 approach into a single-stage recipe that trains discrete and continuous actions together:

1. Train the action expert with flow matching for continuous-action inference.
2. Train the backbone with discrete FAST action tokens, vision-language data and robot-planning data.
3. Stop expert gradients at the backbone keys and values. The expert attends to the image/language/state prefix, not the FAST action tokens; the backbone does not attend to the expert.

In the paper’s frozen-backbone ablation, shirt-folding performance is 0%. FAST supplies a robot-learning signal to adapt the backbone without updates from the action-expert loss. For the generalist table-bussing comparison, π0 needs 7.5 times as many training steps to reach similar performance. These results describe π0.5 + KI and the tested configurations, not every model carrying the π0.5 name Driess 2025↓Knowledge Insulating Vision-Language-Action Models: Train Fast, Run Fast, Generalize BetterDanny Driess, Jost Tobias Springenberg, Brian Ichter et al., 2025.

The language-following evidence also has a qualification: vision-language co-training can give the joint-training baseline good language following without stop-gradient. The authors report improved but still imperfect language following, not complete immunity to interference Driess 2025↓Knowledge Insulating Vision-Language-Action Models: Train Fast, Run Fast, Generalize BetterDanny Driess, Jost Tobias Springenberg, Brian Ichter et al., 2025.

π0.6 and π*0.6: scaling and experience

The π0.6 model card is dated November 17, 2025 Physical Intelligence 2025↓π0.6 Model CardPhysical Intelligence, 2025. It initializes the backbone from Gemma3 4B; the architecture figure labels SigLIP 400M, and the action expert has approximately 860M parameters with the backbone’s layer count. Pre-training accepts up to four 448×448 images. Knowledge Insulation supervises the backbone with FAST tokens while the expert predicts continuous actions without sending gradients into it. With five denoising steps and three camera inputs, chunk inference takes 63 ms on one H100. In the card’s out-of-the-box evaluation without task-specific fine-tuning, π0.6 folds laundry reliably and fully assembles the box 20% of the time; the card says these tasks previously required fine-tuning to achieve non-zero success Physical Intelligence 2025↓π0.6 Model CardPhysical Intelligence, 2025.

π*0.6 uses Recap to learn from demonstrations, expert teleoperation corrections during rollouts, and autonomous experience. It begins with offline-RL pre-training, then task-specific demonstration fine-tuning and one or more rounds of collecting experience and retraining Amin 2025↓π*0.6: a VLA That Learns From ExperienceAli Amin, Raichelle Aniceto, Ashwin Balakrishna et al., 2025.

Its language-conditioned distributional value function predicts negative remaining steps for successful episodes and a large negative value for failures, normalized per task. Actions receive a binary improvement indicator from an advantage estimate and a task-dependent threshold; human corrections are forced positive. The policy learns from both good and bad data, and the default evaluation samples with the positive indicator. This is not simply a test of whether raw value increased Amin 2025↓π*0.6: a VLA That Learns From ExperienceAli Amin, Raichelle Aniceto, Ashwin Balakrishna et al., 2025. The companion blog illustrates credit assignment with a bad portafilter grasp that causes a later insertion failure; it does not give a measured twenty-second delay Amin 2025↓π*0.6: a VLA that Learns from ExperienceAli Amin, Raichelle Aniceto, Ashwin Balakrishna et al., 2025.

The report evaluates a static bimanual setup with two 6-DoF arms, parallel-jaw grippers, three cameras and 50 Hz joint-position control. Compared with offline RL plus task-specific demonstration fine-tuning, adding on-robot experience more than doubles successful completions per hour for double-shot espresso and the diverse-laundry evaluation. The espresso trial must complete all steps within 200 seconds without critical mistakes. Diverse-laundry training covers 11 item types, but the quantitative test uses a challenging button-up shirt, folded and stacked within 500 seconds Amin 2025↓π*0.6: a VLA That Learns From ExperienceAli Amin, Raichelle Aniceto, Ashwin Balakrishna et al., 2025.

Keep the report's qualifications with those results. Section VI-C describes failure reduction as “about a factor of two,” whereas Figure 8 says “more than 2×.” Its “90%+” success summary excludes diverse laundry. The easier laundry task is a T-shirt or shorts folded and stacked within 200 seconds; the box task has a 600-second limit. The box chart reports separate stages, and Section VI-C.2 describes about 90% success for folding and labeling while noting remaining failures, especially final placement. This does not establish strictly greater than 90% end-to-end success for every application Amin 2025↓π*0.6: a VLA That Learns From ExperienceAli Amin, Raichelle Aniceto, Ashwin Balakrishna et al., 2025.

The MEM report evaluates π0.6-MEM with two complementary memory scales. Its video encoder interleaves bidirectional spatial attention with causal-temporal attention across observations, while a high-level policy predicts updated natural-language summaries of earlier semantic events. Those summary predictions are trained using labels generated by an off-the-shelf language model; they are not an external agent writing the runtime notes Torne 2026↓MEM: Multi-Scale Embodied Memory for Vision Language Action ModelsMarcel Torne, Karl Pertsch, Homer Walke et al., 2026.

The authors report tasks requiring up to fifteen minutes of memory, including recipe setup and kitchen cleanup, not a fifteen-minute dense-video input. Their separate adaptation experiments train on targeted corrections: adjusting chopstick grasp height after misses and changing which way a refrigerator door is opened. Both memory and memoryless policies receive correction data. Chopstick evaluation uses the lowest table-height setting after collecting data in the upper half of the height range; refrigerator success requires opening within four grasps. The reported in-context adaptation is therefore conditioned on this training and evaluation setup, not a claim of untrained, indefinite continual learning Torne 2026↓MEM: Multi-Scale Embodied Memory for Vision Language Action ModelsMarcel Torne, Karl Pertsch, Homer Walke et al., 2026.

These sections cite the named model card and lab reports. Those document formats do not establish whether model weights are downloadable or what their licenses permit.

π0.7: generalization as a prompting problem

π0.7 (April 2026) reframes the generalization problem: the bottleneck is underspecified prompts, not model capacity Ai 2026↓π0.7: a Steerable Generalist Robotic Foundation Model with Emergent CapabilitiesBo Ai, Ali Amin, Raichelle Aniceto et al., 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. During training, π0.7 simulates Real-Time Chunkingreal-time chunkingAn inference-time scheme that keeps chunked policies responsive when the model is slower than the control loop. While the robot finishes the current chunk, the policy conditions its next prediction on the actions that will already have executed by the time inference returns, holding that committed prefix fixed and generating only the remainder. Black, Galliker, and Levine introduced it for flow-matching policies, where it tracks moving targets under inference delays that break both naive chunk execution and temporal ensembling. delays of 0 to 12 timesteps, which the paper describes as corresponding to a maximum inference latency of 240 ms on a 50 Hz robot. This is a training setup, not a common latency benchmark; UR5e experiments run at 20 Hz while the other tested robots run at 50 Hz Ai 2026↓π0.7: a Steerable Generalist Robotic Foundation Model with Emergent CapabilitiesBo Ai, Ali Amin, Raichelle Aniceto et al., 2026.

The paper expands the conditioning context while training the model to use different subsets. The dropout rules have an explicit exception: the control-mode label is not dropped Ai 2026↓π0.7: a Steerable Generalist Robotic Foundation Model with Emergent CapabilitiesBo Ai, Ali Amin, Raichelle Aniceto et al., 2026.

• Task and subtask language; the subtask comes from a learned high-level policy or a human supervisor at runtime.
• Episode metadata for overall quality, overall speed and mistakes. The runtime configuration sets quality to 5, mistakes to false and speed from a task-specific episode-length percentile.
• A control-mode label for joint or end-effector control, always retained during training.
• Multi-view subgoal images from a BAGEL-initialized world model. Subgoal and subtask generation run in separate threads; VLA inference uses the latest available context rather than waiting for both.

The paper reports laundry-folding transfer to a bimanual UR5e with no laundry-folding training data on that embodiment. It describes task progress matching experienced teleoperators attempting this task on the UR5e for the first time, with improved performance when visual subgoals are used Ai 2026↓π0.7: a Steerable Generalist Robotic Foundation Model with Emergent CapabilitiesBo Ai, Ali Amin, Raichelle Aniceto et al., 2026. In the air-fryer example, a zero-shot prompt produces only a partial attempt. Repeated language coaching supplies instructions used to fine-tune a high-level policy that then generates the subtasks autonomously, without additional teleoperation Ai 2026↓π0.7: a Steerable Model with Emergent CapabilitiesBo Ai, Ali Amin, Raichelle Aniceto et al., 2026. Read the compositional claim with the authors' own hedge: they describe "early signs of compositional task generalization," and the zero-shot air fryer attempt does not finish the task without coaching Ai 2026↓π0.7: a Steerable Model with Emergent CapabilitiesBo Ai, Ali Amin, Raichelle Aniceto et al., 2026. This capability account does not establish checkpoint availability or licensing.

Which checkpoints are listed

The pinned openpi README lists base checkpoints and download locations for π0, π0-FAST and π0.5. This is a catalogue observed in the retained September 7, 2026 retrieval, not a complete inventory of every later model or a statement about an unlisted model’s license Physical Intelligence 2026↓openpiPhysical Intelligence, 2026. Select a generation in the timeline to inspect its separately named primary source.

Across these named reports, the research questions shift from continuous-action generation and tokenization to heterogeneous co-training, Knowledge Insulation, learning from experience, memory and richer prompts. Those contributions are model- and experiment-specific; repository membership alone does not determine which capabilities a checkpoint has.