π₀: A Vision-Language-Action Flow Model for General Robot Control
1. Paper overview
In one sentence: A VLM and a smaller flow-matching action expert turn diverse robot pre-training into reusable control, but dexterous task mastery still depends on post-training and execution design. ev-overviewev-flowev-architectureev-recipeev-runtimeev-complex-discrepancy
| At a glance | What to know |
|---|---|
| Research problem | Source description Semantic knowledge from images and text does not itself provide precise robot control. The paper targets a reusable policy that handles different embodiments, language instructions and dexterous manipulation, while combining diverse demonstrations with curated task-specific behavior. Its central challenge is to preserve broad competence while generating high-frequency, multimodal continuous actions. ev-overviewev-flowev-recipe |
| Core mechanism | Source description A PaliGemma-initialized VLA adds a roughly 300M-parameter action expert to a 3B VLM, yielding about 3.3B parameters. The experts exchange information through self-attention inside one transformer. ev-flowev-architecture |
| A key reported result | Complex manipulation after post-training: The fine-tuned model exceeds 0.5 on every displayed task, as stated in the caption. Average normalized task progress. Seven tasks in Figure 13, including laundry, box building and egg packing; ten trials per task with task-specific partial-credit rubrics. Usually above out-of-box and scratch variants; egg packing instead favors scratch in the plot. Supports task-dependent benefits of the two-stage recipe, not a greater-than-50% full-task success rate. Table bussing's pre-training membership conflicts between text and panel heading. ev-complexev-complex-discrepancyev-complex-rubric |
| Reading caution | Source description The authors leave dataset composition, weighting, reliable performance and transfer to substantially different domains unresolved. Diverse robot coverage alone does not establish universal control. ev-limitations |
Core contributions
- Source description
A PaliGemma-initialized VLA adds a roughly 300M-parameter action expert to a 3B VLM, yielding about 3.3B parameters. The experts exchange information through self-attention inside one transformer. ev-flowev-architecture
- Reader analysis
The empirical contribution combines large-scale robot pre-training, direct prompting, language guidance and downstream fine-tuning. The authors argue that diverse pre-training supplies recoveries while curated post-training supplies fluent strategies; that causal explanation is stronger than the experiments directly isolate. ev-recipeev-complexev-limitations
Figure 3. The VLM and action expert share a transformer but specialize in different token types. Original paper, p. 4 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read upward from the camera views and the “fold shirt” prompt into the blue VLM region. The green region receives robot state q_t and noisy action tokens; its repeated loops represent the iterative action-generation computation. The left-hand data sources describe training, while the right-hand robot drawings illustrate the embodiments served by the policy. The drawing is an overview rather than a complete attention diagram. Appendix B specifies one transformer with two expert weight sets: action tokens may attend to observations and to each other, but observation tokens cannot attend to noisy actions. This directional restriction is what permits prefix caching during flow integration. ev-overviewev-flowev-architectureev-horizon-discrepancyev-language
What it supports. The smaller action expert provides a continuous control interface to a pre-trained VLM while keeping repeated flow computation confined to the action suffix. The source describes a roughly 3B VLM plus 300M action expert. These components generate robot actions, rather than a predicted visual future or an inverse-dynamics reconstruction.
Where the evidence stops. The drawing ends its action sequence at a_{t+H}, whereas Section IV and Appendix B define H tokens ending at a_{t+H−1}. Use the formal definition, H = 50. A separate high-level policy is an optional deployment component, not shown here.
2. Motivation
2.1 The problem and the proposed response
Semantic knowledge from images and text does not itself provide precise robot control. The paper targets a reusable policy that handles different embodiments, language instructions and dexterous manipulation, while combining diverse demonstrations with curated task-specific behavior. Its central challenge is to preserve broad competence while generating high-frequency, multimodal continuous actions. ev-overviewev-flowev-recipe
2.2 What this reading follows
To understand π0, follow an observation all the way to an executed movement. Images and language enter a pre-trained VLM, while robot state and noisy actions use a smaller expert inside the same transformer. Flow matching converts noise into an action chunk; the controller executes only part of that chunk before looking again. A second distinction matters just as much: broad robot pre-training and curated task post-training play different roles. The six source visuals below connect this architecture to physical experiments and runtime measurements. They also expose why progress scores, task-specific guidance and inconsistent source labels require careful interpretation. ev-overviewev-flowev-architectureev-recipeev-runtimeev-complex-discrepancy
3. Research context
We place the paper in the collection through its world–action interface. The catalog labels and the reading’s assessment are shown separately.
| Catalog dimension | Recorded classification |
|---|---|
| Major category | VLA |
| Architecture | Dual-system |
| Prediction paradigm | Not applicable |
| Quadrant | Not applicable |
3.1 Evidence-based assessment
Conflicts with the recorded classification
VLA and diffusion/flow-matching labels are supported. The unconditional Dual-system architecture label conflicts with Appendix B's explicit single transformer with two expert weight sets. An additional high-level VLM creates a hierarchy only in selected deployments. Neither future-observation prediction nor inverse dynamics is the action mechanism, supporting Not applicable for the world/action prediction quadrant. This assessment does not reclassify the core policy as a One Model world-action model. ev-flowev-architectureev-language
This is the collection’s architectural analysis, not a new related-work survey. Benchmark comparisons and their protocols appear in Section 6.
4. Problem formulation
4.1 Inputs and outputs
| Inputs | Outputs |
|---|---|
|
|
4.2 Equations and their role
5. Method in detail
5.1 1. Follow information through one transformer
Begin with the camera images, the instruction and the measured robot state. Appendix B divides these inputs and the noisy action chunk into three attention blocks: images plus language, state, then actions. Every token may read its own block, but earlier blocks cannot read later ones. Thus the action tokens can combine semantics and proprioception while coordinating the whole proposed motion. Separate expert weights handle the two token families, yet self-attention connects them inside one transformer. This matters for the catalog assessment: two expert parameter sets do not by themselves establish two independently operating policies. A hierarchy arises when a separate high-level VLM supplies intermediate commands in the language experiments. Reader interpretation: the architecture separates computational responsibilities while preserving a shared attention interface, whereas the optional high-level policy separates task selection from movement generation. ev-architectureev-flowev-language
5.2 2. Separate flow time, action horizon and feedback
Training starts from a demonstrated 50-action chunk and Gaussian noise. Their linear interpolation defines a noisy chunk at flow time τ; the network learns the vector pointing from the noise toward the demonstration. At deployment, ten Euler updates integrate this field from noise to actions. Those updates are internal computations, not ten interactions with the physical world. The resulting horizon is also different from the execution interval: the controller applies 16 actions on the 20 Hz systems or 25 on the 50 Hz systems before observing again. The observation prefix can remain cached within one generation because its attention mask excludes the changing action suffix. Reader interpretation: Table I measures the cost of generating a motion proposal, while the open-loop prefix determines how long that proposal proceeds without a fresh observation. Both affect responsive control. ev-flowev-architectureev-runtime
5.3 3. Ask what pre-training actually improves
The authors propose that diverse demonstrations teach broad behavior and recovery, while curated post-training teaches efficient task execution. Figures 11 and 13 support benefits from combining those stages, but the strength of the conclusion depends on the comparison. Within Figure 11, π0 scratch removes robot pre-training while retaining the VLM starting point described in Section VI-C. Comparing these two curves is more targeted than comparing against π0-small, which changes parameter count, visual encoders and the attention design as well. Even the targeted comparison varies by task: the drawer curves nearly coincide with more data, while towel folding improves substantially. Figure 13 adds complex behaviors, yet its egg-packing ordering contradicts an across-the-board advantage. Reader interpretation: the evidence supports useful, uneven transfer; it does not isolate recovery learning or prove that every task benefits. ev-recipeev-finetuningev-baselineev-complexev-complex-discrepancy
5.4 Training and inference
During training
The paper reports about 10,000 hours of in-house demonstrations across seven robot configurations and 68 broadly defined tasks, with 903M timesteps. Open-source data account for 9.1% of the weighted mixture. Section III says the entire OXE dataset is used, but Figure 4 specifies an OXE Magic Soup subset; the exact selection remains unresolved. ev-dataev-mixture-discrepancy
The VLM is further trained on robot data and the new action expert is initialized from scratch. The loss matches a denoising vector field; a shifted beta schedule emphasizes noisy actions and stops at s = 0.999. Task names and approximately two-second segment annotations provide language supervision. No frozen-backbone training stage is specified. ev-overviewev-flowev-sampling
The full base model uses 700k updates. Post-training specializes the model with curated task data; small-data experiments use 1, 5 and 10 hours, while the recipe discusses 100 or more hours for complex tasks. Optimizer, batch size, training hardware and detailed curation thresholds are not specified in the supplied paper. ev-base-protocolev-recipeev-finetuningev-reproduction-gaps
During inference
Starting from Gaussian noise, ten forward Euler steps generate a chunk, with observation keys and values cached. UR5e/Franka execute 16 actions at 20 Hz before re-inference every 0.8 seconds; other robots execute 25 at 50 Hz and re-infer every 0.5 seconds. Chunks execute open-loop between updates; temporal ensembling was tried and rejected. ev-flowev-runtime
Some tasks use a separate high-level VLM to issue intermediate language commands. Human-guided, high-level-model-guided and flat-command evaluations are distinct conditions. Language planning is optional; the low-level policy still generates and executes continuous actions. ev-language
5.5 Implementation flow
- Encode the observation
Image encoders and language tokens enter the VLM expert. Linear projections introduce proprioception, while an MLP combines noisy actions with a sinusoidal flow-time embedding. The model predicts actions conditioned on the current observation, without predicting future images or a world-state rollout. ev-flowev-architecture
- Separate weights and constrain information flow
Image/language tokens use the larger expert; state/action tokens use the smaller expert. Three attention blocks contain images plus text, state, and noisy actions. Attention is bidirectional within each block but cannot reach later blocks; actions therefore access all observations without feeding noise back into the cached prefix. ev-architecture
- Represent multiple embodiments
State and action vectors are padded to the reported maximum dimension of 18, and missing image slots are masked. Sampling weights scale as n^0.43 for each task–robot combination, where n is its sample count, reducing the dominance of larger datasets. ev-data
6. Experiments & results
π0 adapts a pre-trained vision-language model to continuous robot control through a smaller flow-matching action expert. Diverse robot pre-training supplies a reusable policy, while task-specific post-training improves dexterous execution. Real-robot experiments support this recipe, but performance depends on task and training conditions; partial-credit scores and several source inconsistencies limit broad claims of mastery.
6.1 Read the original evidence
Figure 7. Direct prompting performs strongly on tasks represented in robot pre-training. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. First identify the filled dark-blue π0 bars and the white, blue-outlined parity bars. They use the same policy design with different update counts: 700k and 160k. The lighter-blue bars belong to π0-small, while purple and green denote OpenVLA variants and Octo. Each group names both a task and its robot. The vertical axis is average task progress over ten trials. Appendix E makes shirt folding binary, but scores other tasks through object placement or substeps. Accordingly, the first group can be read as success frequency under its flattened-shirt protocol; the remaining groups generally measure partial completion. ev-base-protocolev-base-resultsev-base-rubricev-baseline
What it supports. For shirt folding, the plotted means are 1.0 for full π0, 0.9 for parity and 0.5 for π0-small. Full π0 leads all five groups, and parity also exceeds the other baselines. This establishes useful direct control after robot pre-training under the evaluated conditions.
Where the evidence stops. These are pre-training tasks, not a held-out-task benchmark. Update-count parity does not guarantee equal compute or epochs. π0-small changes architecture and size as well as initialization, and the figure supplies no uncertainty intervals.
Figure 9. Understanding intermediate commands and benefiting from them are separate measurements. Original paper, p. 9 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. The left panel measures whether the policy follows language, comparing π0 with π0-small under expert instructions. The right panel asks whether that capability improves task progress. Solid bars denote flat task commands, diagonal patterns human intermediate commands, and horizontal stripes the high-level VLM. Darker colors belong to π0, lighter colors to π0-small. Compare conditions within the same task before comparing across tasks, because the object counts and instruction sequences differ. Table setting uses seven objects and approximately 20 instructions per episode. The human condition supplies expert guidance; only the flat and high-level-model conditions are autonomous in this comparison. ev-languageev-language-rubricev-baseline
What it supports. Table-setting progress rises from approximately 0.38 with a flat command to 0.83 with high-level VLM guidance; human intermediate commands reach approximately 0.72. These are visual estimates. The larger policy can use intermediate language effectively, while the smaller policy does not show a consistent task-level benefit.
Where the evidence stops. High-level guidance is not uniformly better: the grocery-bagging HL bar is below π0-flat. Model size and other architecture changes confound the VLM-pretraining comparison. Ten trials and absent uncertainty intervals limit the precision of these differences.
Table I. Generating an action chunk takes 73 ms on-board in the reported RTX 4090 setup. Original paper, p. 16 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the three computation rows before the totals. Image encoders consume 14 ms and the observation forward pass 32 ms. The row labeled “x10 action forward pass” is 27 ms for all ten flow iterations together, not for each iteration. Their sum gives 73 ms on-board inference. The additional 13 ms network row produces the 86 ms off-board total. Appendix D specifies three camera images and an NVIDIA GeForce RTX 4090; the mobile robot uses off-board inference over Wi-Fi. The table concerns chunk generation, while execution timing is specified separately in the surrounding text. ev-runtimeev-architectureev-flowev-reproduction-gaps
What it supports. Caching observations and repeatedly evaluating only the action suffix makes iterative flow generation practical in the reported controller. A 50-action proposal does not force the robot to execute all 50: it refreshes after 16 actions on 20 Hz systems or 25 on 50 Hz systems.
Where the evidence stops. The 50 Hz number describes action execution, not fresh camera-conditioned inference. The table supplies no timing distribution or software-version configuration. These hardware-specific measurements do not establish the latency of an independently reproduced implementation.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| Directly prompted shirt folding Pre-training task on Bi-ARX, no post-training; five initially flattened shirts, two trials each, up to approximately five minutes per trial. | 1.0 for full π0 (Figure 7 bar at the 1.0 tick). Mean binary success, plotted as average task progress | π0 parity: 0.9; π0-small: 0.5, read from the plotted bars. Full π0 has 700k updates; parity has 160k. Strong performance on this narrow protocol. It does not measure crumpled-laundry folding or unseen-task success. No uncertainty intervals are supplied. ev-base-protocolev-base-resultsev-base-rubric |
| Language-guided table setting Fine-tuned language evaluation; ten trials, seven objects and approximately 20 instructions per episode. | Approximately 0.83 with high-level VLM commands, visually estimated from Figure 9. Average task progress | Approximately 0.38 with the flat task command and 0.72 with human intermediate commands; plot estimates, not tabulated values. High-level guidance helps this task. The grocery-bagging bars do not show a universal autonomous-guidance benefit; model size and architecture also confound the π0-small comparison. ev-languageev-language-rubricev-baseline |
| Towel folding with five hours of fine-tuning data Bi-ARX downstream task, similar to pre-training shirt folding; ten trials per method/data condition. | 1.0 for pre-trained π0, as plotted in Figure 11. Normalized progress over two folds and final neatness | Approximately 0.8 for π0 scratch and 0.1 for Diffusion Policy. Scratch here removes robot pre-training; the π0 architecture retains its VLM initialization. Robot pre-training helps this condition, but gains vary: drawer results nearly coincide at larger data sizes and the microwave curve is nonmonotonic. ev-finetuningev-newtask-rubric |
| Complex manipulation after post-training Seven tasks in Figure 13, including laundry, box building and egg packing; ten trials per task with task-specific partial-credit rubrics. | The fine-tuned model exceeds 0.5 on every displayed task, as stated in the caption. Average normalized task progress | Usually above out-of-box and scratch variants; egg packing instead favors scratch in the plot. Supports task-dependent benefits of the two-stage recipe, not a greater-than-50% full-task success rate. Table bussing's pre-training membership conflicts between text and panel heading. ev-complexev-complex-discrepancyev-complex-rubric |
| Action-chunk inference latency Three camera images, NVIDIA GeForce RTX 4090, ten flow steps; off-board mobile inference uses Wi-Fi. | 73 ms on-board; 86 ms off-board. Reported latency in milliseconds | Image encoding 14 ms, observation pass 32 ms, all ten action passes 27 ms; off-board network latency adds 13 ms. Measured computation supports chunked control, but these numbers do not imply a fresh visual decision at every 50 Hz actuator step. ev-runtime |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Figure 11. Robot pre-training helps many data-limited tasks, with substantial variation across tasks. Original paper, p. 10 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Follow the solid and dashed blue curves first. Section VI-C distinguishes robot-pretrained π0 from π0 trained on target data without robot pre-training; its discussion retains VLM initialization as the architectural starting point. The horizontal axis counts hours of target-task demonstrations, and the vertical axis is normalized progress. The upper-left panel averages tasks; the remaining panels reveal what that mean conceals. ACT and Diffusion Policy learn from target data, while the OpenVLA and Octo markers use public OXE checkpoints. Their sparse points reflect limited baseline evaluation, not a complete sweep. Use Appendix E's task-specific rubrics when interpreting a value near one. ev-finetuningev-newtask-rubricev-recipe
What it supports. At five hours, towel folding reaches the plotted maximum of 1.0 for π0, versus roughly 0.8 for scratch and 0.1 for Diffusion Policy. The drawer curves nearly coincide at larger data sizes, while microwave performance is nonmonotonic. Transfer therefore changes both with the task and with available demonstrations.
Where the evidence stops. Baselines have different pre-training histories and incomplete data-size coverage. Ten trials per condition provide limited precision, with no error bars. The ablation supports robot-pretraining benefits in selected conditions but does not isolate recovery learning or guarantee monotonic improvement.
Figure 13. The two-stage training recipe improves most complex tasks, with an explicit exception. Original paper, p. 11 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. The solid bars show the full robot-pretraining-plus-fine-tuning recipe, diagonal bars training on task data without robot pre-training, and outlined bars the base policy without task post-training. Read each task group as a comparison of these recipes, not as a probability of completing a long demonstration. Appendix E awards partial credit: laundry includes extraction, flattening, folding and placing or stacking; egg packing includes individual eggs and lid closure. The panel titles divide tasks by purported pre-training membership. That split is not fully reliable: the table-bussing variant is described as absent from pre-training on page 10 but is placed in the upper panel. ev-complexev-complex-discrepancyev-complex-rubric
What it supports. The caption's defensible numerical summary is that the fine-tuned model exceeds 0.5 normalized progress on every displayed task. Laundry and box building show substantial gains over the alternatives. Egg packing is an exception to a universal advantage: its scratch bar is higher than the fine-tuned bar.
Where the evidence stops. Preserve two source conflicts: table bussing's panel membership contradicts its text description, and egg packing contradicts the prose claim of being best across the board. The caption's “typically” is narrower. Partial-credit scores do not establish full-episode success rates.
7. Analysis & limitations
7.1 What the evidence leaves open
The authors leave dataset composition, weighting, reliable performance and transfer to substantially different domains unresolved. Diverse robot coverage alone does not establish universal control. ev-limitations
The base comparison's update-count parity does not establish equal FLOPs or epochs. Fine-tuning compares π0's in-house pre-training with public OXE checkpoints and task-only ACT/DP. π0-small also changes size, encoders and attention architecture, so it cannot isolate VLM initialization alone. ev-base-protocolev-finetuningev-baseline
Figure 3 ends the chunk at a_{t+H}, while the formal definition ends at a_{t+H−1}. Figure 13's table-bussing category and egg-packing ordering conflict with nearby prose. These discrepancies remain explicit rather than being silently reconciled. ev-horizon-discrepancyev-complex-discrepancy
7.2 Questions for discussion
- Does diverse robot pre-training improve recovery specifically, or primarily improve common action and perception patterns?
- When does intermediate language guidance help after controlling for the extra information supplied?
- How much does task success depend on the open-loop execution interval rather than action-chunk generation quality?
8. Reproducibility audit
8.1 Requirements and known gaps
Reproduction requires compatible multi-camera robots, PaliGemma initialization, cross-embodiment demonstrations, language segment labels and curated task data. The paper does not provide a complete training configuration, precise dataset split manifest, action normalization specification or high-level policy configuration; access to the large in-house dataset is not established by the PDF. ev-dataev-recipeev-architectureev-languageev-reproduction-gaps
Reader-proposed minimum checks: vary robot pre-training while holding architecture and target data fixed, then test cached action generation against uncached evaluation with identical noise. Record task progress, complete success and latency separately using Appendix E's rubrics. These are proposed checks, not reproduced findings. ev-finetuningev-architectureev-runtimeev-newtask-rubric
8.2 Proposed reproduction checks
The following checks are proposals motivated by the paper. They have not been run as part of this reading.
Check 1: Does robot pre-training specifically improve recovery?
Reader-proposed experiment, not performed: compare mixed-robot-pretrained π0 with the same π0 architecture initialized only from the same VLM, using identical one-hour paper-towel-replacement data, post-training update budgets and evaluation objects. Run matched clean episodes and episodes with a standardized displacement of the replacement roll before the next observation. Report the four-point Appendix E score, complete success, and recovery time separately across repeated training seeds. A larger pre-training benefit after displacement than in clean trials would support the recovery explanation. Similar clean and perturbed gains, or no reproducible gain, would weaken that specific mechanism even if average performance improves. Exact data membership and training settings would need to be documented because the PDF does not fully specify them. ev-recipeev-finetuningev-newtask-rubricev-reproduction-gaps
Check 2: Does prefix caching preserve the action computation?
Reader-proposed implementation check, not performed: use identical observations, weights, initial Gaussian noise, numerical precision and ten Euler steps in cached and fully recomputed inference. Compare the vector field at every step and the final 50-action chunk within a predeclared numerical tolerance. Vary the noisy action suffix while holding observations fixed and verify that prefix keys and values remain unchanged, as Appendix B's mask requires. Then time image encoding, observation processing and all ten action passes separately on an RTX 4090 with three images, excluding network latency before adding it as a separate measurement. A material output difference would expose an incorrect mask or cache; identical outputs with lower repeated-prefix cost would support the claimed optimization. ev-flowev-architectureev-runtime
8.3 Reading coverage
Visual audit: Original pages 1–12 and 15–17 were rendered and visually inspected, including the title/authors/version/affiliation, Figures 1–14, Table I, equations, method and training descriptions, comparison protocols, limitations, architecture appendix and every retained scoring rubric. Pages 13–14 contain references and were read in the complete eight-chunk text pass. All six final PNG crops were individually viewed; labels, axes, legends and table rows are readable. Figure 3's horizon label was checked against pp. 5 and 15; Figure 13's categories and ordering against pp. 10–11; the Gaussian sampling discrepancy against p. 5. No separate supplements, linked videos, code or experiments were inspected.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 15, 16, 17. Appendix coverage: reviewed.
Original figures and tables remain the work of the source’s authors. Extractions preserve their scientific content; any HTML wrapper layout is disclosed with each figure. The surrounding reading notes are our own.
Text reading scope & known omissions
- Abstract and title page (p. 1)
- I. Introduction (pp. 2–3)
- II. Related Work (pp. 3–4)
- III. Overview (p. 4)
- IV. The π0 Model (pp. 4–5)
- V. Data Collection and Training Recipe, A–C (pp. 5–6)
- VI. Experimental Evaluation, A–D (pp. 6–11)
- VII. Discussion, Limitations, and Future Work (pp. 11–12)
- Acknowledgements and References (pp. 12–15)
- Appendix A. Contributions (p. 15)
- Appendix B. Model Architecture Details (pp. 15–16)
- Appendix C. Non-VLM Baseline Architecture (p. 16)
- Appendix D. Inference (p. 16)
- Appendix E. Evaluation Details, A–D (pp. 16–17)
Outside the original text pass
- Identity/version disclosure: the supplied primary PDF is arXiv:2410.24164v4, dated 8 January 2026 on p. 1. The catalog identifies RSS 2025. This report reads the supplied v4 preprint, not an independently compared proceedings edition; equivalence of their scientific content is not established here.
- The observed title π0: A Vision-Language-Action Flow Model for General Robot Control matches the catalog's π₀ rendering. The inspected title page lists 24 authors, including Michael Equi, and omits catalog-listed Laura Smith. The catalog uses Michael Robert Equi; Laura Smith appears in the PDF acknowledgements on p. 12. The supplied catalog remains unchanged.
- Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
- Separate supplemental material availability has not been fully verified.
- Some figure-label fonts extract as /C... glyph codes; inspect the retained PDF visually for these figures and labels.
- The extraction limitations above were addressed for this report by inspecting original page renders and six final crops. All eight supplied chunks were individually read. Reference-only pages 13–14 were read as text; all other pages were also visually inspected.
- No external links, videos, code, datasets or separate supplements were inspected, and no experiments were reproduced. Implementation and release availability are not verified.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
ev-identityPDF p. 1, title, author block, affiliation footnote and arXiv margin; p. 12, Acknowledgements
Title and 24 named authors are printed on p. 1, including Michael Equi and excluding Laura Smith; Physical Intelligence is located in San Francisco, California, USA. The margin identifies arXiv:2410.24164v4, 8 January 2026. Laura Smith is acknowledged on p. 12.
Go to primary source ↓ev-overviewPDF pp. 2–3, Section I; p. 4, Section III and Figure 3 caption
The framework combines VLM initialization, cross-embodiment robot pre-training and task-specific post-training; language labels include task names and roughly two-second segment annotations.
Go to primary source ↓ev-flowPDF p. 5, Section IV, action-chunk definition, unnumbered flow-matching loss, sampling construction and Euler equation
The policy models p(A_t|o_t) with H=50, images, language and state; trains a vector field toward A_t−ε; generates actions with ten Euler steps of size 0.1; initializes a 300M action expert from scratch alongside a 3B PaliGemma VLM.
Go to primary source ↓ev-architecturePDF p. 15, Appendix B, Additional inputs and outputs; Incorporating the flow matching timestep; Attention mask; Action expert
One transformer has two expert weight sets interacting through self-attention. Image/text tokens use the VLM expert, state/actions the action expert. Images/text, state and actions form three blockwise causal attention groups, each internally bidirectional. Action outputs alone decode to the vector field; state/observation keys and values can be cached.
Go to primary source ↓ev-samplingPDF pp. 15–16, Appendix B, Sampling the flow matching timestep; Figure 14
The shifted beta distribution emphasizes lower, noisier flow times and excludes times above s=0.999. This is motivated by the difficulty of predicting conditional mean actions; no isolated schedule ablation is reported here.
Go to primary source ↓ev-dataPDF p. 3, Section II, demonstration scale; pp. 5–6, Section V-A; p. 6, Section V-C and Figure 5
The paper describes about 10,000 demonstration hours, 903M in-house timesteps, seven robot configurations and 68 tasks, with 9.1% weighted open-source data. Task–robot mixtures use n^0.43 weights; vectors are padded to 18 dimensions and missing camera slots masked.
Go to primary source ↓ev-recipePDF pp. 5–6, Section V and V-A, pre-training/post-training discussion
Pre-training seeks behavioral diversity and recovery coverage; curated post-training seeks consistent, fluent execution. The general recipe describes task-specific datasets ranging from five hours to 100 or more hours.
Go to primary source ↓ev-mixture-discrepancyPDF p. 4, Section III, first paragraph; p. 5, Figure 4 caption
Section III calls OXE entire, whereas Figure 4 explicitly calls the used OXE mixture a subset named OXE Magic Soup. The source does not resolve this selection difference.
Go to primary source ↓ev-base-protocolPDF p. 7, Section VI-A, baseline training and evaluation metric paragraphs
Full π0 uses 700k updates and parity π0 160k; OpenVLA has 160k and Octo 320k. The base evaluation uses the same robot-data mixture, plus a UR5e-only OpenVLA comparison, with ten episodes per task and normalized partial progress except binary shirt folding.
Go to primary source ↓ev-base-resultsPDF p. 8, Figure 7, Shirt Folding group and legend; adjoining Section VI-A discussion
Shirt-folding bars show full π0 at 1.0, parity at 0.9 and π0-small at 0.5. Full π0 is highest across all five displayed tasks, and parity exceeds the other baselines. The plot supplies no error bars or numerical data table.
Go to primary source ↓ev-base-rubricPDF p. 17, Appendix E-A, Shirt folding and other base task rubrics
Five flattened shirts each receive two binary-scored trials capped at 15,000 steps, approximately five minutes. Success requires sleeves folded inward and a lengthwise half-fold. Other tasks score sorted objects or manipulation substeps.
Go to primary source ↓ev-languagePDF p. 6, Section V-B; p. 8, Section VI-B; p. 9, Figure 9, both panels and caption
Flat commands, expert human intermediate commands and autonomous high-level VLM commands are separate conditions averaged over ten trials. Table-setting task progress is approximately 0.38 flat, 0.72 human and 0.83 HL by visual reading. The grocery-bagging HL bar is below the flat bar. π0-small differs in size and initialization.
Go to primary source ↓ev-language-rubricPDF p. 17, Appendix E-B, language instruction following
Language evaluation scores correct object repositioning and instruction following. Table setting has seven objects and about 20 instructions per episode; bussing and bagging use different episode structures.
Go to primary source ↓ev-finetuningPDF p. 9, Section VI-C, tasks, baselines and last paragraphs; p. 10, Figure 11, legend and five-hour Towel Folding point
Five downstream tasks use 1/5/10-hour data conditions and ten trials. Robot-pre-trained π0 is compared with π0 without robot pre-training, task-only ACT/DP and public OXE-pretrained OpenVLA/Octo. The text distinguishes π0's VLM initialization from robot pre-training. Towel folding at five hours reaches 1.0 versus approximately 0.8 scratch and 0.1 DP. Other task curves vary and are not uniformly monotonic.
Go to primary source ↓ev-newtask-rubricPDF p. 17, Appendix E-C, new dexterous task rubrics
Towel folding awards three points: first fold, second fold and neatness. Paper-towel replacement awards four: grasp/remove old roll and grasp/place new roll. Other tasks have their own partial-credit denominators.
Go to primary source ↓ev-complexPDF pp. 10–11, Section VI-D, tasks and comparison; p. 11, Figure 13 and caption
Seven complex tasks compare fine-tuned π0, scratch and out-of-box variants over ten trials. The caption states the fine-tuned model exceeds half the maximum score on all tasks and typically outperforms ablations. The plots show substantially stronger fine-tuned laundry and box-building results.
Go to primary source ↓ev-complex-discrepancyPDF p. 10, Section VI-D, Table bussing description; p. 11, Section VI-D results paragraph and Figure 13 panel headings/Packing Eggs group
The complex table-bussing description says it is absent from pre-training, but Figure 13 places it under tasks present in pre-training. The prose claims the full recipe is best across the board, but the egg-packing scratch bar is higher than the fine-tuned bar. The caption's typically outperforms phrasing is narrower.
Go to primary source ↓ev-complex-rubricPDF p. 17, Appendix E-D, complex task evaluation rubrics
Laundry scores extraction, flattening, folding and placing/stacking per garment; two trials for each of five garments are capped at about five minutes. Box building awards five substep points and egg packing seven. These are partial-credit progress scores, not episode success fractions.
Go to primary source ↓ev-baselinePDF p. 5, Non-VLM baseline paragraph; p. 8, Sections VI-A/VI-B confound discussion; p. 16, Appendix C
π0-small has about 470M parameters, DistilBERT language features, smaller pre-trained image encoders with unshared weights, an observation encoder and a cross-attending DiT action expert. Its differences from π0 extend beyond VLM initialization; it still uses pre-trained visual encoders.
Go to primary source ↓ev-runtimePDF p. 16, Appendix D and Table I, all timing rows
With three images on RTX 4090, encoders take 14 ms, observation pass 32 ms, ten action passes together 27 ms; totals are 73 ms on-board and 86 ms with 13 ms network latency. The controller executes 16 actions/0.8 s at 20 Hz or 25 actions/0.5 s at 50 Hz, without temporal ensembling.
Go to primary source ↓ev-limitationsPDF pp. 11–12, Section VII, especially p. 12 limitations paragraph
The authors leave mixture composition and weighting, reliable task mastery, required data quantities and transfer across robots/tasks unresolved; extrapolation to driving, navigation and legged locomotion is future work.
Go to primary source ↓ev-equation-discrepancyPDF p. 5, Section IV, adjacent Gaussian path and noisy-action construction
The printed path is N(τA_t,(1−τ)I), while the explicit sample is τA_t+(1−τ)ε for ε∼N(0,I). Under standard covariance notation these differ by a square in the covariance; no correction is supplied.
Go to primary source ↓ev-horizon-discrepancyPDF p. 4, Figure 3, action output labels; p. 5, Section IV, A_t definition; p. 15, Appendix B, Additional inputs and outputs
Figure 3 labels its last action a_{t+H}; the formal definitions use a_{t+H−1} and explicitly specify H=50 action tokens. The report uses the formal definition.
Go to primary source ↓ev-reproduction-gapsPDF pp. 5–6, Section V-A/B; p. 9, Section VI-C; pp. 15–16, Appendices B–D
The source supplies mixture weighting, data amounts, architecture and inference settings but no complete optimizer/batch/training-hardware configuration, precise split manifest, detailed curation thresholds, action normalization specification or high-level policy configuration. These sections do not establish release/access conditions for the in-house dataset.
Go to primary source ↓8.5 Primary sources
π0: A Vision-Language-Action Flow Model for General Robot Control ↗
PDF · 13,539 extracted words
Source fingerprint
fbdfba56258bdbd220207ba63410c6c943e43a73e97a7a5a47ca0bc74204f821