π_0.7: a Steerable Generalist Robotic Foundation Model with Emergent Capabilities
1. Paper overview
In one sentence: Detailed context lets π0.7 learn from heterogeneous robot experience and accept generated visual goals, improving steerability while adding annotation, runtime-compute and evaluation burdens. e02e04e05e11e15e17e18
| At a glance | What to know |
|---|---|
| Research problem | Author claim Diverse robot trajectories contain conflicting strategies, failures and different execution speeds. The authors argue that merely pooling them encourages unsuitable behavior averaging. They ask whether richer context can make mixed-quality robot and non-robot data useful while retaining precise instruction following and supporting new skill combinations. e02 |
| Core mechanism | |
| A key reported result | Zero-shot UR5e shirt folding: π0.7 (GC): 85.6% progress and 80% success. Task progress; full-task success rate. Static-bimanual folding data transferred to bimanual UR5e without target-robot folding demonstrations; flattened shirt start and joint control. Ten experienced human operators performed three trials each. Human operators: 90.9% progress and 80.6% success, as reported. Comparable observed performance, not a statistical equivalence test. Success requires the full six-point rubric. The reported human percentage cannot be reconstructed as a simple binary count over 30 trials; aggregation is unexplained. e15e16e23 |
| Reading caution | Author claim The authors report lower reliability for unseen tasks or task–robot combinations and explicitly cannot certify that all supposedly novel skills are absent from their broad datasets. Related behaviors can occur incidentally or under other labels. e19 |
Core contributions
Figure 2. Language, visual goals and execution preferences meet in the action policy. Original paper, p. 4 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start at the lower left: either the human or the learned high-level policy produces a semantic subtask. Follow that branch both upward into the VLA prompt and across to the world model. The world model also receives current observations and metadata, then sends generated subgoal images upward. The upper VLA consumes those goals alongside observation memory, task language and metadata. At the far right, noise enters the action expert and continuous commands emerge. Figure 2 is a runtime information-flow overview; the separate training description explains FAST supervision and the stop of action-expert gradients into the backbone. e03e04e05e09e10e21e22
What it supports. The world model provides an intermediate visual specification that can transfer knowledge into control. Its output is not itself a robot action: execution still requires the VLA and the physical controller. This separation supports the catalog’s dual-system interpretation when generated goals are enabled.
Where the evidence stops. The approximately 5B parameter count refers to the VLA, not the full deployed stack. The separate world model is 14B and adds substantial inference cost. Generated goals are optional; this diagram does not imply that every reported run uses them.
2. Motivation
2.1 The problem and the proposed response
Diverse robot trajectories contain conflicting strategies, failures and different execution speeds. The authors argue that merely pooling them encourages unsuitable behavior averaging. They ask whether richer context can make mixed-quality robot and non-robot data useful while retaining precise instruction following and supporting new skill combinations. e02
2.2 What this reading follows
A command such as “fold the shirt” leaves many choices unspecified: where to grasp, which fold to make, and how quickly to work. π0.7 expands the context around each training trajectory so that one policy can represent different strategies and qualities of execution. At deployment, desired metadata selects a behavior mode, while a separate world model can supply visual subgoals. The paper’s most useful evidence connects this mechanism to physical outcomes and data ablations. Read the specialist comparisons, cross-robot tests and coached tasks as distinct protocols: strong results in one do not establish unattended generalization in all three. e02e04e05e11e15e17e18
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 | WAMs |
| Architecture | Dual-system |
| Prediction paradigm | IDM |
| Quadrant | Q4 · Dual-system × IDM |
3.1 Evidence-based assessment
Supports the recorded classification
The recorded Dual-system × IDM classification is supported for the goal-conditioned configuration: a separate world model predicts visual subgoals, and a VLA infers actions conditioned on current observations and those goals. This is architectural evidence, not merely joint training. It does not describe every runtime mode: π0.7 can omit goals and act as a language-conditioned VLA. The paper does not present a single shared model jointly emitting future images and actions, nor action-conditioned rollout search. e03e04e05e09e21
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 Turn mixed-quality experience into selectable behavior
Imagine two successful laundry episodes: one finishes cleanly and quickly, while the other repeatedly misses a grasp. A short task label assigns both the same context even though their action sequences differ. π0.7 adds episode length, quality and mistake labels, so training can associate different trajectories with different conditions. At runtime the prompt requests high quality, no mistakes and a short task-specific episode length. Dropout also teaches the policy to operate with missing metadata, enabling classifier-free guidance. Reader interpretation: this is conditional behavior selection, not an explicit guarantee that the robot will achieve the requested score. Figure 7 tests removing metadata, and Figure 18 asks whether it changes the value of adding lower-quality data. Both are stronger evidence for the proposal than simply observing a successful rollout. e05e06e09e12e18
Figure 19. Attention masks separate action supervision and guidance branches, while exposing an ordering ambiguity. Original paper, p. 22 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Follow the labels in each panel rather than treating the shaded shape as a generic causal triangle. The upper panels distinguish observation memory, text, optional image goals, flow actions and FAST action tokens. The caption explicitly states that FAST and flow-action tokens do not attend to one another; FAST is training-only. The upper-right panel packs positive and negative metadata-guidance branches that remain isolated. Below, the goal generator receives observation encodings from both ViT and VAE, with image goals in the VAE representation. Its inference panel separates text/image guidance into three groups. These masks concern information access; gradient insulation is specified separately in Section III. e03e07e20e21
What it supports. The implementation uses different masks for training and guided inference, so a single generic causal mask would miss key distinctions. In particular, branch isolation is necessary for the packed computations to represent separate conditional predictions. The goal generator and action policy also have different conditioning structures.
Where the evidence stops. Section VI-B on page 6 places causal text after the observation/goal blocks, but this figure and its caption place image goals after text. Both locations were inspected. The intended goal/text ordering remains unresolved; this crop faithfully preserves the appendix version.
5.2 Use a predicted image as a control target, then close the loop
Language identifies the next semantic step, but a multi-view goal can also depict the intended grasp and object configuration. The BAGEL-based world model generates this intermediate target from current observations, subtask language and metadata. The VLA then predicts an action chunk conditioned on observation history and the available prompt. Its training includes both actual future frames and generated goals, addressing the mismatch between perfect demonstrations and imperfect generated images. At runtime, the robot continues acting while new goals are computed; goals refresh after a subtask change or four seconds. Only part of each action chunk executes before another prediction. This is a feedback controller using generated visual context, not a guarantee that the predicted future will occur, and not a search over action-conditioned simulated rollouts. e04e08e09e10e21e22
5.3 Separate transferred control skill from supplied task decisions
The paper asks several different generalization questions. In the specialist comparison, tasks were represented in training, and the achievement is strong performance without downstream specialist fine-tuning. In UR5e folding, the missing data is the target-robot/task combination: the policy must adapt a known skill to another morphology. In new long kitchen tasks, a human initially chooses the next subtask through coaching. Training a high-level policy on those coaching episodes later automates those decisions without collecting new low-level demonstrations. Reader interpretation: these results form a useful progression, but they cannot be collapsed into one claim of zero-shot autonomy. The authors also acknowledge that related skills and appliances can occur in the broad training mixture. A convincing reproduction must document both the assistance supplied at evaluation and the exposure excluded during training. e06e11e15e16e17e19
5.4 Training and inference
During training
The mixture includes demonstrations, failures, autonomous evaluations, interventions, public robot data, egocentric human video and auxiliary web tasks. Generalization-focused evaluation rollouts are excluded. Knowledge insulation trains the backbone with FAST-token cross-entropy while stopping action-expert gradients into it; this does not freeze the backbone. e03e06
Goals appear in 25% of training examples; subtask language is dropped in 30% of goal-conditioned examples. Metadata is dropped entirely with probability 15%, plus 5% independently per component; control mode is retained. Quality ranges from 1 to 5; episode length uses 500-step bins. e05
Real goal sampling uses segment endpoints with probability 0.25 and uniformly sampled futures 0–4 seconds ahead otherwise. Generated goals also train the VLA to reduce mismatch. World-model supervision uses segment-end images and high-quality temporal language labels. RTC training simulates 0–12-step delays. e08e07e21
During inference
Runtime metadata requests quality 5, mistake=false and the task-specific 15th percentile episode length. Goals refresh on a subtask change or after four seconds, asynchronously. Five denoising steps produce each 50-step chunk; 15 or 25 steps are executed before replanning with RTC. e09
Metadata classifier-free guidance contrasts conditioned and metadata-dropped predictions, using reported weights 1.3, 1.7 or 2.2. Human coaching is an inference input. Later training a high-level language policy on coaching episodes yields autonomy without new teleoperated low-level demonstrations; it is still additional training. e09e17
5.5 Implementation flow
- Encode observations and context
The roughly 5B VLA combines a Gemma3 4B backbone with an 860M flow-matching action expert. Its MEM encoder compresses up to six history frames per view into a single-frame token budget. It accepts up to four observation views and three goal views, resized to 448×448; proprioceptive history receives linear embeddings. e03e07
- Specify the next achievable scene
A human or high-level policy supplies the next subtask. The separate 14B BAGEL-based model conditions on current views, subtask and metadata to generate near-future multi-view images. These specify object arrangements and gripper configurations; they are conditioning targets, not verified physical outcomes. e04e21
- Convert goals into physical control
The action expert attends to VLA context and denoises continuous commands. With future images present, the authors characterize action prediction as inverse dynamics. A PD controller executes joint targets; end-effector targets first undergo numerical inverse kinematics. e05e07e10
6. Experiments & results
π0.7 turns a generalist robot policy into a steerable controller by conditioning on detailed subtasks, desired episode quality and speed, and optional generated subgoal images. A separate world model supplies visual goals; the VLA predicts executable action chunks. The evidence supports strong performance on trained tasks and selected transfers, but novel long tasks can require coaching, and broad training exposure makes novelty difficult to certify. This report concerns the supplied April 24, 2026 v2.
The supplied paper presents its empirical comparisons as plots in Figures 6–18 and 20–22, not as a numbered quantitative results table. This edition therefore uses original result and ablation plots among its six crops. No table is reconstructed from bar heights. Most plots omit exact numerical values, trial counts and error-bar definitions, limiting quantitative precision; the explicit shirt-folding values are retained from the accompanying text. e11e12e14e16e18
6.1 Read the original evidence
Figure 6. A single generalist approaches specialist performance on tasks represented in training. Original paper, p. 9 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the two rows as different comparisons. In the upper row, each task has a throughput panel and a success panel. Throughput means successful episodes per hour and is normalized to that task’s RL specialist, whose reference level is one. Success uses its own percentage scale. The lower row changes both comparator and metric: the grey policies are SFT specialists, and the vertical axis is partial task progress. Photographs identify the physical setup; they are not additional evaluations. Compare yellow and grey within the same panel rather than pooling heights across tasks or interpreting progress as full completion. e06e11
What it supports. The strongest visual conclusion is breadth: the same generalist reaches broadly similar performance across several dexterous tasks. Diverse laundry and box building show a throughput advantage over the RL specialist reference, while other panels are closer. This is evidence of useful generalist training and specialist-behavior distillation.
Where the evidence stops. These are trained tasks, and π0.7 training includes prior-policy experience. Out-of-the-box means no task-specific post-training here. Exact bar values and the meaning of the drawn error bars are not supplied, so the figure does not establish statistical equivalence.
Figure 12. Transfer tests change the robot as well as the required manipulation strategy. Original paper, p. 12 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. First inspect the robot icons and arrows beneath each task: transfer directions differ, so this is not one pooled benchmark. The left group covers rearrangement and bagging, including the larger morphology change for shirt bagging. The right group transfers towel and shirt folding from a smaller static bimanual robot to bimanual UR5e. Solid yellow denotes π0.7; hatched yellow adds generated goals. The blue dashed human reference applies to shirt-folding task progress, not to success rate or every task. Appendix F supplies the human protocol and Figure 22 separately compares progress and full-task success. e15e16e23
What it supports. Generated goals improve the plotted folding transfer, supporting their role as useful control context. For shirt folding, the accompanying text reports 85.6% progress and 80% success for π0.7 (GC), versus 90.9% and 80.6% for humans. These are physical execution outcomes on the target robot.
Where the evidence stops. No target-robot folding demonstrations does not mean the target robot or folding skill was absent from all training. The human study includes three trials per operator without prior practice; its success aggregation and the policy trial count are unspecified. Comparable values do not establish equivalence.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| Zero-shot UR5e shirt folding Static-bimanual folding data transferred to bimanual UR5e without target-robot folding demonstrations; flattened shirt start and joint control. Ten experienced human operators performed three trials each. | π0.7 (GC): 85.6% progress and 80% success. Task progress; full-task success rate | Human operators: 90.9% progress and 80.6% success, as reported. Comparable observed performance, not a statistical equivalence test. Success requires the full six-point rubric. The reported human percentage cannot be reconstructed as a simple binary count over 30 trials; aggregation is unexplained. e15e16e23 |
| Out-of-the-box dexterity on trained tasks A single generalist evaluated against task-specific π0.6 RL or SFT specialists on physical robots. | Figure 6 shows broadly comparable performance, with higher throughput on diverse laundry and box building; exact bar values are not tabulated. Success rate and specialist-normalized throughput; task progress for the SFT comparisons | Separate task-specific specialists, rather than another identically trained generalist. No task-specific post-training for π0.7 does not mean the tasks or specialist-generated data were unseen. e06e11 |
| Metadata and autonomous-data ablations Laundry, espresso and box-building tasks; remove metadata or autonomous evaluation data. | Both ablations underperform the full recipe across the plotted tasks, especially in throughput. Success and throughput normalized to full π0.7 | π0.7 (no metadata) and π0.7 (no eval data). Supports both ingredients; removing evaluation data also changes the available training trajectories. e12 |
| Instruction following in unseen environments Four unseen kitchens and two unseen bedrooms; 14 scenarios with 3–6 instructions each. Separate complex-reference and reversed-bias tasks. | π0.7 improves over π0.5/π0.6; generated goals help complex references and the reversed microwave-to-fridge task. Instruction-following success; task progress for reversed-bias tasks | Prior policies and π0.7 without generated goals. Distinct protocols are not pooled. Goal conditioning is task-dependent, not uniformly beneficial. e14 |
| Coaching new long-horizon tasks Air-fryer loading/unloading and bagel toasting; no task-specific action-level training episodes. Later high-level training uses coaching episodes. | π0.7 follows coaching more effectively than prior models; learned high-level prompting approaches coached performance. Task progress | π0.5/π0.6 and live-coached versus autonomous π0.7. The initial result includes human subtask selection. The subsequent autonomous result includes high-level training. e17 |
| Learning from mixed-quality and diverse data Laundry models use top 30%, 50%, 80% or all data, with/without metadata; separate generalization ablation removes the most diverse or random 20%. | Metadata yields the best throughput with all data; removing the most diverse 20% hurts more than removing random 20%. Laundry successes/hour; unseen-task success rate | Eight laundry models and equal-quantity diversity-removal controls. Supports context-dependent gains from broader data. The plotted laundry curve is not strictly increasing at every intermediate point; no universal scaling law follows. e18 |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Figure 7. Metadata and autonomous experience each contribute to the full training recipe. Original paper, p. 9 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. The legend has three conditions: full π0.7, no metadata, and no evaluation data. For each task, inspect the throughput panel before the adjacent success panel. The normalization has changed from Figure 6: full π0.7 is now the throughput reference at one. The other two bars therefore measure losses relative to the full recipe. Box building makes the separation particularly clear, while simple laundry shows a smaller success-rate separation than its throughput separation. The no-evaluation-data condition removes autonomous episodes, including the route for inheriting strong prior policies; the no-metadata condition removes information used to distinguish episode quality and strategy. e05e06e11e12
What it supports. The plots support the claim that heterogeneous experience becomes more useful when its quality is identifiable. Both removals hurt, and throughput often separates conditions more clearly than success alone. The result motivates examining how the policy selects efficient behavior rather than evaluating only whether an episode eventually finishes.
Where the evidence stops. Removing autonomous data changes the training trajectories as well as their quality distribution. This figure is not a complete factorial interaction test of metadata and data diversity, and its error bars do not provide an explicitly defined uncertainty estimate.
Figure 18. More data helps differently depending on its context and task diversity. Original paper, p. 14 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. The left panel uses a quality-ranked laundry dataset. Moving from 30% to 100% includes progressively lower-quality episodes; the two curves distinguish training with and without episode metadata. Its vertical axis is raw successes per hour, unlike the normalized throughput in Figures 6 and 7. The right panel is a separate experiment: compare the full model, random removal of 20%, and removal of the most diverse 20% on three short-horizon unseen tasks. The two removal conditions retain equal quantities of data. This control asks whether the removed examples’ content matters beyond the simple loss of training volume. e05e11e12e18
What it supports. The full-data metadata model has the highest laundry throughput, whereas the no-metadata curve falls when the lowest-quality data is added. Removing highly diverse examples also causes larger generalization losses than random removal. Together these experiments support conditioning and diversity as complementary parts of the recipe.
Where the evidence stops. The metadata curve is not strictly increasing between every intermediate subset, despite the caption’s broad language about continuous improvement. Diversity removal controls data quantity, but it does not isolate every semantic property of the removed examples or establish a universal scaling law.
7. Analysis & limitations
7.1 What the evidence leaves open
The authors report lower reliability for unseen tasks or task–robot combinations and explicitly cannot certify that all supposedly novel skills are absent from their broad datasets. Related behaviors can occur incidentally or under other labels. e19
Most result plots lack exact values, trial counts and an explicit definition of their error bars. They support qualitative comparisons, but not reconstructed confidence intervals or significance claims. The human study provides more protocol detail, yet leaves aggregation unclear. e11e12e14e16e18
The main text implies goal-image tokens precede text; Figure 19 and its caption put goals after the text prompt. This unresolved attention-order discrepancy matters for implementation. Separately, the world-model objective maximizes a quantity called a flow-matching loss without explaining its sign convention. e20e04
7.2 Questions for discussion
- Does correctly assigned quality metadata outperform equally informative but shuffled labels under matched training budgets?
- How much transfer survives when generated goals are stale or visually plausible but incompatible with the target robot?
- Can a documented exposure audit distinguish new skill composition from incidental task coverage?
8. Reproducibility audit
8.1 Requirements and known gaps
A faithful retraining needs the actual data mixture, annotation process and model checkpoints. The source describes modalities and dropout but does not give total training-data scale, mixture weights, optimizer schedule, training steps or total training compute. e06e05e21
Appendix D reports VLA inference on an H100: 38 ms for a minimal configuration and up to 127 ms with memory/goals. The separate goal generator uses four H100s, 8-bit matrix multiplications and modified SageAttention: 25 denoising steps take 1.25 s. These are inference measurements, not a training budget. e22
Match 20 Hz control on UR5e and 50 Hz on other robots, camera placement, command representation and asynchronous execution. For shirt folding, obtain the referenced final-state quality chart and numerical time limit: the supplied rubric names the chart, but neither item is fully specified here. e10e16e23
Proposed checks: compare real versus shuffled quality metadata at fixed data/compute; separately ablate valid versus mismatched generated goals while logging goal age and physical success. These test conditioning mechanisms without treating image plausibility as task completion. e05e08e09e12e18
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 metadata carry useful information beyond an extra prompt field?
Reader-proposed experiment, not performed: use the same laundry trajectories, model initialization, optimizer budget and goal-conditioning schedule to train three conditions with correct metadata, metadata shuffled between episodes within the same task, or metadata omitted. Preserve the marginal label distribution in the shuffled condition. Evaluate the same held-out physical starts using identical high-quality prompts and fixed guidance settings. Record success, successes/hour and scored mistakes, with uncertainty over independent runs. Correct labels should outperform shuffled labels if meaningful quality/strategy disambiguation drives the benefit. If shuffling preserves the gain, the claimed information mechanism is weakened. Repeat on a high-quality subset and the full mixed-quality set to test the interaction. e05e09e12e18
Check 2: Do generated goals help through grounding, and when does latency erase that help?
Reader-proposed experiment, not performed: hold the VLA, metadata, language prompts, initial scenes and joint controller fixed on UR5e shirt folding and the reversed microwave-to-fridge task. Compare no goal, the current generated goal, and a goal mismatched to the scene but drawn from the same task. Then inject controlled delays into the valid-goal condition while retaining asynchronous execution. Log goal age at every action update and score physical progress and full completion separately. A grounding explanation predicts valid goals outperform mismatched goals; comparable performance would weaken it. Increasing delays should reveal any stale-goal failure regime. Obtain the missing shirt-quality chart and declare scoring/time limits before evaluating. e08e09e10e14e15e16e22e23
8.3 Reading coverage
Visual audit: Visually inspected the title/author/version page, all numbered Figures 1–22, Eq. (1), the world-model objective, Algorithm 1, training/inference descriptions, result protocols, limitations, hardware details and task rubrics on the declared pages. All six final original crops were separately inspected; Figure 19 was rendered at higher resolution to preserve its small labels. Figure 2 arrows were checked against Sections V–VII; Figure 19 was checked against Section VI-B and its caption, revealing the disclosed goal/text-order conflict. Figure 18 was checked against Section IX-E; the intermediate metadata points are not strictly monotonic. The complete 25-page text, including references and Appendix A–G, was read. Separate supplements, code and underlying experimental data were not supplied for inspection.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 22, 23, 24, 25. 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
- Title, author block and abstract (pp. 1–2)
- I. Introduction (p. 2)
- II. Related Work (pp. 2–3)
- III. Flow-Based Vision-Language-Action Models (p. 3)
- IV. π0.7 Overview (pp. 3–4)
- V. Diversifying the Prompt, A–E (pp. 4–5)
- VI. The π0.7 Model and Training Recipe, A–C (pp. 5–7)
- VII. Prompting π0.7 at Runtime, including Algorithm 1 (p. 7)
- VIII. Robot System Details (pp. 7–8)
- IX. Experimental Evaluation, A–E (pp. 8–15)
- X. Discussion and acknowledgements (p. 15)
- References (pp. 15–21)
- Appendix A–G: contributions, attention, world-model training, inference, action spaces, human study, and task rubrics (pp. 22–25)
Outside the original text pass
- Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
- The preceding acquisition limitation was addressed by visually inspecting PDF pages 1–15 and 22–25, including all numbered figures. References on pages 16–21 were read as text.
- Separate supplemental material availability has not been fully verified.
- Only arXiv:2604.15483v2, dated 24 April 2026, was supplied. The catalog submission date is 16 April 2026; earlier versions were not supplied, so scientific revision differences cannot be established. The catalog title uses π_0.7 where the observed title uses π0.7; the work and listed authors match.
- Code, project websites, underlying datasets and external references were not inspected; no experiments were reproduced.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e01PDF p. 1, title/author block and arXiv margin
The observed title is π0.7: a Steerable Generalist Robotic Foundation Model with Emergent Capabilities. Physical Intelligence and the complete individual author list match the catalog. The margin identifies arXiv:2604.15483v2 [cs.LG], 24 Apr 2026.
Go to primary source ↓e02PDF pp. 2–3, Sections I–II
The authors motivate richer context as a way to disambiguate heterogeneous behavior and enable compositional generalization; they emphasize a training methodology rather than a newly proposed architecture.
Go to primary source ↓e03PDF pp. 3–4, Section III, Eq. (1), Section IV and Figure 2 caption
Observation history and prompt condition action chunks. FAST cross-entropy trains the VLM; action-expert gradients are stopped at the backbone. The VLA uses Gemma3 4B, an included 400M vision encoder and an 860M action expert, approximately 5B total.
Go to primary source ↓e04PDF pp. 4–5, Sections V-A/B, Figure 2 and unnumbered world-model objective
Human/high-level subtasks and a separate BAGEL-initialized 14B model provide language and multi-view visual goals. The world-model equation prints maximization of L_CFM, subsequently called the standard flow-matching loss; the sign convention is not explained.
Go to primary source ↓e05PDF p. 5, Sections V-C/D/E
Episode metadata describes length in 500-step bins, quality 1–5 and mistakes. Joint/end-effector mode is a text identifier. Goals occur in 25% of examples, associated language dropout is 30%, whole-metadata dropout 15%, component dropout 5%; control mode is not dropped. Goal-conditioned prediction is described as inverse dynamics.
Go to primary source ↓e06PDF pp. 5–6, Section VI-A and footnote 1
Training includes demonstrations, suboptimal/autonomous episodes, interventions, public robot data, human video and web tasks. Prior RL-policy data contributes to distillation. Generalization-focused evaluation rollouts are excluded. Total mixture proportions and full training schedule are not provided.
Go to primary source ↓e07PDF pp. 6–7, Section VI-B
Up to four camera views, six history frames per view, three goal views, 448×448 resizing and one-second history stride are specified. State is linearly embedded. The expert uses 50 bidirectionally attending action tokens and adaptive RMSNorm; RTC simulates 0–12-step delays, up to 240 ms at 50 Hz.
Go to primary source ↓e08PDF p. 7, Section VI-C
For real goals, 25% are segment-end frames and 75% are sampled uniformly 0–4 seconds ahead. Additional generated-goal training examples address mismatch with runtime images; the real/generated mixture ratio is not specified.
Go to primary source ↓e09PDF p. 7, Section VII, Algorithm 1 and CFG expression
Desired metadata uses task-specific 15th-percentile length, quality 5 and no mistakes. Goals regenerate after a subtask change or four seconds; all components run asynchronously. Five denoising steps produce 50-step chunks, of which 15 or 25 steps execute. Metadata CFG uses beta in {1.3, 1.7, 2.2}.
Go to primary source ↓e10PDF pp. 7–8, Section VIII and Figure 4
The platforms include mobile and static bimanual robots and UR5e systems with parallel-jaw grippers. UR5e runs at 20 Hz and others at 50 Hz. A PD controller applies actions; numerical inverse kinematics converts end-effector targets to joint positions.
Go to primary source ↓e11PDF pp. 8–9, Section IX-A and Figure 6; p. 24, Appendix G, Laundry/Make Espresso/Box Building rubrics
The dexterity comparison concerns trained tasks without π0.7 task-specific post-training. Figure 6 compares a single generalist to RL and SFT specialists; its top row uses success and throughput normalized to the RL specialist, while the lower row uses task progress. Laundry and box building show throughput advantages; precise plotted values and error-bar definitions are not tabulated.
Go to primary source ↓e12PDF p. 9, Figure 7 and Section IX-A ablation paragraph
Removing metadata or autonomous evaluation episodes reduces reported performance. Throughput is normalized to full π0.7, unlike Figure 6. The strongest visible gaps include box building; no exact-value table accompanies the plots.
Go to primary source ↓e13PDF p. 10, Figure 8 and Section IX-A memory paragraph; p. 24, Appendix G, memory-task rubrics
The generalist compares favorably to task-specific memory specialists on Swap 3 Mugs, Find Object, Scoop Coffee and Window Cleaning. Appendix G calls the coffee task Scoop Beans and specifies two scoops.
Go to primary source ↓e14PDF pp. 10–11, Section IX-B and Figures 9–11
The novel-environment evaluation spans four kitchens, two bedrooms and 14 scenarios of 3–6 instructions. Complex-reference comparisons use success; reversal tasks use progress. Generated goals help complex references and reverse fridge-to-microwave behavior, but do not uniformly improve every condition.
Go to primary source ↓e15PDF pp. 11–12, Section IX-C and Figures 12–13; p. 23, Appendix E and Figure 20
Transfer directions and target morphologies vary by task. Laundry transfers from smaller static bimanual robots to bimanual UR5e without UR5e folding data; goal conditioning improves the plotted folding results. Qualitative examples show changed grasp strategies. Main transfer tests use joint control; prior-model EE comparisons show no clear overall benefit.
Go to primary source ↓e16PDF p. 13, Section IX-C continuation; pp. 23–24, Appendix F and Figures 21–22
Ten operators from the top 2% by experience, averaging about 375 hours, performed three trials each without practice. Start state, time limit and criteria matched policy evaluation. The source reports human progress/success of 90.9%/80.6% and π0.7 (GC) of 85.6%/80%. Trial counts for the policy and the aggregation producing the human success percentage are not specified.
Go to primary source ↓e17PDF pp. 13–14, Section IX-D and Figures 14–17
Short tasks can run from direct prompts. Novel longer kitchen tasks use stepwise human coaching. Coaching episodes subsequently train a high-level language policy without collecting new low-level teleoperation. Similar appliances occur in human/external data; direct task-specific robot action examples were not collected.
Go to primary source ↓e18PDF pp. 14–15, Section IX-E and Figure 18
Laundry data is ranked by quality/speed into top 30%, 50%, 80% and 100% subsets, training eight models with/without metadata. Full-data metadata training performs best; intermediate plotted points are not strictly monotonic. Removing the highest-diversity 20% degrades unseen-task performance more than random 20% removal.
Go to primary source ↓e19PDF p. 15, Section X, limitation paragraphs
Authors describe weaker zero-shot reliability than on seen tasks and acknowledge that broad datasets make truly unseen tasks hard to certify because related skills can appear incidentally or under other labels.
Go to primary source ↓e20PDF p. 6, Section VI-B attention paragraph; p. 22, Appendix B, Figure 19 and caption
The main text describes observation/goal bidirectional blocks followed by causal text. Figure 19 and its caption instead place the goal block after the text prompt. The caption states FAST and flow-action tokens do not attend to each other, and positive/negative CFG branches remain isolated; the world model uses three CFG groups.
Go to primary source ↓e21PDF pp. 5 and 22, Section V-B and Appendix C
The BAGEL world model uses high-quality segmented robot/human-video labels plus image-editing/video data. Three current views and segment-end targets are encoded with ViT and VAE; separate 7B understanding and generation backbones process them. ViT input is 448×336 and VAE input 512×384. Detailed dataset proportions and optimizer/training budget are absent.
Go to primary source ↓e22PDF pp. 22–23, Appendix D
Gemma3-based VLA/high-level inference is described on a single H100. Minimal VLA latency is 38 ms; memory and goals raise the worst case to 127 ms. The 14B goal generator uses four H100s, 4-way tensor parallelism, 8-bit matrix multiplies and modified SageAttention, reaching 1.25 s for 25 denoising steps with text/image CFG. Robot action continues asynchronously.
Go to primary source ↓e23PDF p. 25, Appendix G, Shirt Folding and Reverse Fridge to Microwave rubrics
Shirt folding begins flattened, awards up to three fold points plus 0–3 quality points, and requires all six for success. The referenced final-state quality chart is not reproduced in the supplied PDF. The reversed food-transfer task awards six substeps, including both door operations and plate transfer.
Go to primary source ↓8.5 Primary sources
π0.7: a Steerable Generalist Robotic Foundation Model with Emergent Capabilities ↗
PDF · 19,161 extracted words
Source fingerprint
54843c88fbabf6bce53ab2ee94cafc9a55c3d5639f7b74c5e8ad24485fe6863b