π^*_0.6: a VLA That Learns From Experience
1. Paper overview
In one sentence: RECAP turns experience into binary advantage labels that improve a flow-matching robot policy, while relying on human feedback and a critic whose calibration and data recipe require care. e01e03e04e09e10e11e20e21
| At a glance | What to know |
|---|---|
| Research problem | Source description Imitation alone inherits demonstration speed and deployment mistakes. RECAP asks how a large VLA can learn from heterogeneous experience and sparse task-outcome feedback while retaining expressive continuous action generation. Its learning loop assumes observations serve as Markovian states; this is explicitly a simplification. e01e02 |
| Core mechanism | Source description The contribution is an integrated recipe: distributional value learning, binary advantage conditioning, and repeated collection of autonomous experience with optional corrective teleoperation. It applies reinforcement learning during both generalist pre-training and task specialization. e02e03e04e07 |
| A key reported result | Targeted collar-up folding failure removal: 97%, explicitly reported after two iterations. Strict episode success rate. One orange T-shirt, fixed adversarial flattened start, correctly centered face-up collar and completed fold within 200 seconds. Initial offline RL + SFT policy: approximately 23% in Figure 12. A narrow failure mode improves substantially. The no-correction-data claim is unresolved because Appendix F describes corrections for this ablation. e09e14e21 |
| Reading caution | Source description The system requires human outcome labeling, episode resets, and sometimes interventions. Exploration is largely greedy and depends on policy stochasticity and corrections. Updates occur between batches rather than concurrently with collection. e15 |
Core contributions
- Source description
The contribution is an integrated recipe: distributional value learning, binary advantage conditioning, and repeated collection of autonomous experience with optional corrective teleoperation. It applies reinforcement learning during both generalist pre-training and task specialization. e02e03e04e07
- Reader analysis
The controlled-data laundry comparison supports the authors’ choice of advantage conditioning over their AWR and PPO implementations, but does not establish superiority over every PPO implementation or online data-collection regime. e13
Figure 3. The critic supplies a training condition to the VLA’s action generators. Original paper, p. 4 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start at the lower cameras, prompt, and metadata. The value head estimates return; the middle box combines observed rewards with the difference between later and current values. The upward path then applies the strict inequality A>ε, agreeing with Equation (3), before entering the VLA. At the top, distinguish language outputs, discrete action tokens, and the separate continuous action expert driven by noise. Section V-A says the continuous expert does not consume predicted discrete actions. The caption and Section V-A also specify a stop-gradient from the action expert into the rest of the model: shared conditioning does not mean every loss updates every block. e03e04e05e08e16e20
What it supports. The mechanism learns how to act under a favorable quality label without fitting a future-image generator. Continuous and discrete action supervision coexist, while a separate critic interprets recorded experience. At default inference, requesting the positive label selects the learned conditional policy; the diagram is a training interaction, not an online planning loop.
Where the evidence stops. Section V-B places the indicator after subtask text, but Appendix C’s Equation (6) includes it in the subtask factor. Threshold calibration also conflicts between Section V-D and Appendix F. Follow the explicit sequence and A>ε rule while preserving these unresolved recipe differences.
2. Motivation
2.1 The problem and the proposed response
Imitation alone inherits demonstration speed and deployment mistakes. RECAP asks how a large VLA can learn from heterogeneous experience and sparse task-outcome feedback while retaining expressive continuous action generation. Its learning loop assumes observations serve as Markovian states; this is explicitly a simplification. e01e02
2.2 What this reading follows
A robot can finish a task correctly and still work too slowly to be useful. RECAP addresses both problems by learning from the outcomes of its own attempts. A separate value model estimates progress toward successful completion; its estimates label recorded actions for advantage-conditioned VLA training. The deployed policy then requests the positive-conditioned behavior. The figures below connect that training mechanism to physical laundry, espresso, and packaging experiments. Read throughput alongside success, and keep the narrower evaluation tasks in view. The source reports substantial gains, but conflicting threshold instructions and intervention counts prevent treating the printed recipe as fully reproducible. e01e03e04e09e10e11e20e21
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
Supports the recorded classification
VLA and flow-matching categories are supported. Dual-system is defensible for the separate value learner and policy, and the VLA backbone/action-expert separation. This does not establish a predictive world-model/controller pair: no future-observation model or inverse-dynamics extraction is introduced. The world/action prediction paradigm and quadrant remain not applicable. e03e04e05
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 Convert an outcome into an action-quality condition
Start with the sparse success label, not the positive token. The reward gives every nonterminal step a cost, so a successful episode completed sooner has a better return. A distributional critic learns empirical returns across demonstrations and collected trials. To label a post-training action, RECAP compares the current value with rewards plus the value observed later in the same trajectory. Appendix F sets that lookahead to 50 steps. Thresholding the resulting advantage produces the extra policy input. Corrective human actions bypass that estimate and receive a positive label. Reader interpretation: this makes the critic a bridge from coarse outcomes to local learning signals, but also concentrates risk in value errors and threshold calibration. The contradictory pre-training threshold descriptions should therefore be tested explicitly, not treated as equivalent implementations. e03e04e06e08e20
Figure 4. Value estimates respond to local mistakes and recoveries within real trajectories. Original paper, p. 5 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read each panel horizontally: the frames describe events, while the line below tracks the critic’s value against elapsed seconds. The reward definition on page 7 makes successful progress move toward zero after task-length normalization. Red bands mark setbacks; green bands mark progress. In the laundry example, crumpling the shirt produces a sharp drop, and later recovery raises the estimate. In the fridge example, opening the door raises value even though the overall episode ultimately fails when the filter tips. This is why RECAP needs a local advantage estimate rather than attaching one episode label indiscriminately to every action. e03e04e06e19
What it supports. The examples make the critic’s intended role concrete: distinguish a setback from subsequent recovery and distinguish intermediate progress from final success. Appendix B adds further illustrative traces. Such temporal variation can supply more selective action labels than an episode-wide success tag alone.
Where the evidence stops. These are selected qualitative examples, not a held-out calibration study. The failing fridge episode includes a locally successful door opening; an increase in predicted value is not proof of eventual task completion or of a correctly calibrated advantage threshold.
5.2 Separate learning the positive branch from requesting it
During training, both favorable and unfavorable examples remain useful because the VLA can explain them under different indicator values. Dropping the indicator on some examples additionally teaches an unconditional branch. The architecture combines token supervision with a flow-matching action expert, whose gradients are insulated from the rest of the model. At inference, the ordinary setting requests positive-conditioned actions with β=1; it need not run the critic to rank newly generated action candidates. Subtask text is predicted first and guides lower-level actions. Optional classifier-free guidance sharpens the continuous policy using conditional and unconditional predictions, with the source warning about aggressive behavior at high weights. Reader interpretation: the expensive evaluation of experience has been moved into training, while deployment uses a learned condition to select behavior. Appendix C’s conflicting subtask conditioning notation remains unresolved. e04e05e08e16e17
5.3 Read the intervention claim separately from the performance claim
The experiments answer several distinct questions. Throughput asks how much correct work the robot completes, success asks whether an episode meets task-specific rules, and the strict laundry test asks whether a recurring collar-orientation error can be removed. Its reported 97% success is a performance observation. The stronger claim that this improvement required no correction data depends on data lineage, which the main text and appendix describe differently. Similarly, the extraction comparison holds recorded data fixed but changes the policy-training recipe and advantage-conditioned pre-training. Reader interpretation: the evidence supports useful behavior changes while leaving some mechanism attribution open. A reproduction should retain fresh evaluation episodes, log autonomous and corrected segments separately, and report productive speed alongside success so that an apparent improvement cannot be explained solely by slower, more conservative execution. e09e10e11e13e14e21
5.4 Training and inference
During training
Pre-training uses tens of thousands of demonstration hours across tasks and robots, plus vision-language web data. Discrete outputs use token likelihoods; continuous actions use a flow-matching loss motivated by a likelihood lower bound. The indicator is randomly dropped 30% of the time to learn conditional and unconditional behavior. e03e05e08e16
Appendix F specifies a 50-step lookahead for post-training advantages and a single-value-call empirical-return estimate for pre-training. It targets approximately 30% positive demonstration labels, generally 40% positive post-training rollout data, and 10% for T-shirts/shorts. The main-text threshold description conflicts with these appendix settings. e08e20
During inference
Predict a subtask, then generate flow-matching actions conditioned on that subtask and a positive indicator. Subtask prediction runs less frequently than action generation; joint/gripper commands execute at 50 Hz. The critic supplies training labels rather than an online search or action-selection loop. Default evaluation uses β=1. e04e05e06e10
Optional classifier-free guidance combines conditional and unconditional predictions for β>1. High guidance can push actions toward the learned support boundary and cause aggressive motions; Appendix E gives moderate examples of β between 1.5 and 2.5. e17
5.5 Implementation flow
- Turn outcomes into progress targets
Successful terminal steps receive zero reward, unsuccessful terminal steps a large negative penalty, and other steps −1. Task-length normalization places value predictions on a negative-to-zero scale. A separate language-conditioned value model learns empirical returns using cross-entropy over 201 bins; its expected value supplies the critic. e03e06
- Label relative action quality
Estimate whether the recorded trajectory segment improved expected return relative to its starting observation, then threshold that advantage. Human corrective actions are forced positive. This labels actions in an existing dataset; it does not simulate candidate futures or perform inverse dynamics. e04e08
- Train a policy that can select good behavior
The VLA uses a Gemma 3 4B backbone and an 860M flow-matching action expert; the separate value backbone totals 670M. FAST actions and subtask text provide autoregressive supervision. Knowledge Insulation prevents action-expert gradients from updating the rest of the VLA, while that backbone still learns through token objectives. e05e06
- Aggregate experience and restart fine-tuning
After offline RL pre-training, task demonstrations produce an initial SFT policy with the indicator fixed positive. Collect trials and optional corrections, refit the critic, and fine-tune the VLA on accumulated data. Both models restart from their pre-trained checkpoints each iteration to reduce drift. e07
6. Experiments & results
RECAP improves a flow-matching VLA by using a separately trained value function to label recorded actions as relatively advantageous, then conditioning action generation on the positive label. Demonstrations, autonomous trials, and optional human corrections feed repeated offline updates. Real-robot experiments report faster and more reliable laundry folding, espresso preparation, and box assembly, but several recipe and data-accounting inconsistencies limit exact reproduction.
The supplied paper presents quantitative results and ablations as Figures 7–12 rather than numerical tables; no original quantitative table is available to crop. This edition therefore includes the original result and diagnostic plots. Exact bar values and numerical error magnitudes are not tabulated, so graph-derived numbers are explicitly approximate. Figure 4 supplies qualitative critic diagnostics rather than a quantitative value-calibration experiment. e10e11e12e13e14e19
6.1 Read the original evidence
Figure 7. Experience improves useful work per hour across four physical task settings. Original paper, p. 9 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the vertical label literally: throughput is successful completions per hour, so both failed trials and slow execution matter. Each panel has its own numerical scale. Compare the pale offline-RL-plus-SFT bar with the bright final RECAP bar to examine the addition of deployment experience. The earlier green bars distinguish supervised pre-training from offline RL pre-training. Error bars are standard errors according to the original caption. Page 8 supplies the task boundaries: diverse-laundry numbers concern button-up shirts, espresso concerns double shots, and boxes must be assembled and stacked within the specified trial deadline. These are not interchangeable workloads. e09e10e12
What it supports. Approximate final versus SFT readings are 8.3 versus 3.8 successes/hour for button-up shirts and 29 versus 14.5 for espresso. The corresponding simple-laundry and box values are about 60 versus 33 and 13 versus 10. The strongest throughput gains therefore depend on the task and comparator.
Where the evidence stops. The paper provides plots rather than exact numerical tables. Treat these readings as approximate and compare within panels. Figure 9’s separate iteration narrative uses improvement ratios that do not clearly match its plotted initial and final endpoints.
Figure 8. Success complements throughput, with a separate interpretation required for box subtasks. Original paper, p. 9 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. The first three panels report episode success under the task-specific rules on page 8. Read the fourth differently: every model has separate retrieval, folding, labeling, and stacking bars. The vertical text inside those bars identifies the stage, so no one bar is the complete box workflow. Human raters supply quality judgments that the evaluation aggregates into success labels; the caption identifies standard-error bars. Comparing the same model stages as in Figure 7 shows whether increased throughput coincides with improved reliability. Avoid multiplying or averaging the box-stage bars into an end-to-end rate without a stated conditional-probability protocol. e09e10e11
What it supports. Espresso success rises from roughly 39% after offline RL plus SFT to roughly 92% for the final policy. Diverse-laundry success remains visibly lower than final simple-laundry and espresso success. Faster execution therefore accompanies stronger reliability, but the level achieved and the meaning of each success bar depend on the evaluation.
Where the evidence stops. The caption says each training stage improves performance, yet several plotted transitions reverse, including offline RL pre-training on box subtasks. Preserve those reversals. Stage-wise box success also cannot establish a single end-to-end success probability.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| Laundry folding: diverse items, button-up-shirt evaluation Real bimanual robot; trained across 11 clothing types, evaluated on button-up shirts; fold and stack within 500 seconds. | Approximately 8.3, read from Figure 7. Successful completions per hour | Approximately 3.8 for offline RL + SFT; the authors describe more than doubled throughput. This is difficult-item evaluation, not an 11-type average. Error bars are standard errors; exact plotted values are not tabulated. e09e10 |
| Cafe: double-shot espresso Commercial espresso machine; complete preparation and serving within 200 seconds without critical errors. | Approximately 29 successes/hour and 92% success, read from Figures 7–8. Throughput and episode success rate | Offline RL + SFT: approximately 14.5 successes/hour and 39% success. Both speed and reliability improve in physical execution. These estimates do not quantify every drink type or verify long-duration video claims. e09e10e11 |
| Laundry folding: T-shirts and shorts Variable basket initial conditions; fold and stack within 200 seconds; extraction baselines use the same RECAP-collected data. | Approximately 60, read from Figures 7 and 11. Successful completions per hour | Approximately 33 for offline RL + SFT, 30 for AWR, and 22 for the implemented PPO variant. AWR retains high success but is slower. The comparison also changes advantage-conditioned pre-training; PPO uses an off-policy stabilization recipe. e09e10e13 |
| Box assembly Factory deployment; assemble, label, and stack a flattened box within 600 seconds. | Approximately 13, read from Figure 7. Successful completions per hour | Approximately 10 for offline RL + SFT. Figure 8 reports retrieval, folding, labeling, and stacking success separately; stage rates must not replace end-to-end success probability. e09e10e11 |
| Targeted collar-up folding failure removal One orange T-shirt, fixed adversarial flattened start, correctly centered face-up collar and completed fold within 200 seconds. | 97%, explicitly reported after two iterations. Strict episode success rate | Initial offline RL + SFT policy: approximately 23% in Figure 12. A narrow failure mode improves substantially. The no-correction-data claim is unresolved because Appendix F describes corrections for this ablation. e09e14e21 |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Figure 11. The extraction comparison shows why success alone can hide slower behavior. Original paper, p. 10 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Use the left panel for productive speed and the right for reliability. The final RECAP bar is yellow; AWR and PPO appear to its right. Section VI-C.3 says the extraction methods receive the same data collected using RECAP, which controls one major source of variation. However, Section VI-B starts AWR from π0.6 without advantage conditioning, and Appendix D does the same for PPO. That PPO baseline uses a single-step diffusion likelihood bound and an SPO-style trust-region modification to stabilize the offline setting. Read this as a comparison of these implemented training recipes, not a universal ranking of algorithm families. e09e10e13
What it supports. RECAP reaches approximately 60 successes/hour, compared with about 30 for AWR and 22 for PPO. AWR’s relatively high success alongside much lower throughput illustrates that reliable imitation of the dataset can still be slow. The experiment supports measuring both axes when evaluating policy extraction.
Where the evidence stops. Using identical recorded data does not equalize pre-training or every optimization choice. The PPO result specifically concerns the paper’s off-policy adapted implementation; a different online collection schedule or trust-region implementation was not tested here.
Figure 12. Targeted experience sharply improves a narrowly defined collar-up folding behavior. Original paper, p. 10 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. The two panels use the same sequence: offline RL pre-training, the initial task policy at i=0, then two RECAP iterations. The left axis measures strict success; the right measures successful completions per hour. Read the task definition on page 8 before interpreting the bars: one orange shirt starts flattened in a fixed arrangement, and success requires the collar to face upward. Section VI-C.4 adds that it must be centered and describes the initial condition as adversarial for the SFT policy. The near-top final bar thus concerns removal of this particular mistake, not unrestricted folding generalization. e09e14e21
What it supports. The text explicitly reports 97% strict success after two iterations, consistent with the final bar; the initial SFT bar is approximately 23%. Throughput also rises visibly. This is strong evidence that the trained policy can change a persistent local behavior under a deliberately demanding success definition.
Where the evidence stops. The main text attributes this result to learning without interventions, using 600 trajectories per iteration. Appendix F instead lists about 1,000 autonomous and 280+378 correction episodes for the ablation. The score is reported, but an autonomous-only causal interpretation remains unresolved.
7. Analysis & limitations
7.1 What the evidence leaves open
The system requires human outcome labeling, episode resets, and sometimes interventions. Exploration is largely greedy and depends on policy stochasticity and corrections. Updates occur between batches rather than concurrently with collection. e15
The Monte Carlo critic estimates returns of a changing behavior mixture, not an optimal off-policy Q-function. Positive labels depend on critic and threshold quality; illustrated value traces are qualitative examples, not calibration measurements. e03e19
Figure 9 plots roughly 33→60 laundry and 10→13 box successes/hour from i=0 to i=2. The accompanying claims of 50% and 2× improvements do not follow from those endpoints; their denominators are unclear. Box throughput initially drops. e12
Notation remains inconsistent: Appendix F writes the pre-training return sum from episode start, whereas Equation (1) defines return from the current step. Appendix C conditions the subtask factor on I, whereas Section V-B places I after the subtask. This report follows the explicit main-text sequence without repairing the appendix. e03e05e08e16
7.2 Questions for discussion
- Does advantage conditioning retain its speed advantage when pre-training, fine-tuning data, and compute are all matched?
- How much strict failure removal persists without corrections once contradictory dataset accounts are resolved?
8. Reproducibility audit
8.1 Requirements and known gaps
A faithful run needs pre-training data/checkpoints, outcome annotations, task metadata, and the two-arm, three-camera platform. The PDF lacks complete optimizer schedules, hardware/compute budgets, all task mixture weights, and explicit held-out split accounting sufficient to reconstruct every result. e03e05e06e08e09e18
Resolve threshold calibration: Section V-D says the 30th percentile of predicted values; Appendix F targets 30% positive advantages using a 10,000-point sample. These differ in quantity and tail fraction. Record actual positive-label fractions rather than assume equivalence. e20
Resolve data lineage: the failure study says two batches of 600 trajectories without corrections, but Appendix F lists about 1,000 autonomous and 280+378 correction episodes. Box data are called 600 autonomous trials per iteration in the main text and 600 demonstrations in Appendix F, both with 360 correction episodes. e21
Reader-proposed checks: compare calibrated versus shuffled advantage labels on identical data; compare autonomous-only versus correction-augmented learning on strict collar-up folding with matched budgets and fresh evaluations. These test label informativeness and intervention dependence separately. e04e14e18e20
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 the advantage label carry useful information?
Reader-proposed experiment: start from identical checkpoints and a fixed T-shirts/shorts dataset, holding updates, sampling weights, indicator dropout, and β=1 inference constant. Compare the Appendix F target of roughly 10% positive labels against labels shuffled within each task while preserving that fraction, plus an all-positive control. Log the achieved label fraction and critic errors on held-out trajectories. Evaluate fresh basket starts with the same 200-second deadline, reporting throughput and success with uncertainty. If calibrated labels do not outperform shuffled labels at comparable success, the claim that advantage information drives the speed gain is weakened. Separately record how the conflicting pre-training threshold recipes alter label fractions before attempting a full reproduction. e04e08e09e13e20
Check 2: Can collar-up failure removal succeed without corrections?
Reader-proposed experiment: use the same initial SFT checkpoint and strict shirt setup, then compare autonomous-only collection, collection allowing logged teleoperator corrections, and SFT retraining on the original demonstrations alone. Match the two experience arms on robot-time and training budgets, report human intervention time separately, and evaluate on fresh trials without interventions after each iteration. Keep the main-text 200-second success rule and report both 97%-style strict success and throughput, without treating the published score as a required outcome. If gains appear only in the correction arm, the autonomous-only interpretation is unsupported; if both experience arms improve, quantify the incremental benefit of corrections. This directly addresses the incompatible data counts in Section VI-C.4 and Appendix F. e07e09e14e18e21
8.3 Reading coverage
Visual audit: Visually inspected the title and complete author block on page 1; overview and task images on pages 1–2; preliminaries on page 3; architecture, value loss, advantage equations, sequence, and Algorithm 1 on pages 4–6; reward, robot setup, and training loop on page 7; task rules on page 8; all quantitative and ablation plots on pages 9–10; failure-removal account and limitations on page 11; and Appendix A–F evidence on pages 16–18, including likelihood, PPO, CFG, thresholds, and conflicting dataset counts. Cross-checked the architecture’s upward advantage path and strict inequality against Equations (2)–(3), and gradient/sequence interpretation against the caption and Sections V-A–B. All six final original crops were separately viewed and checked for readable labels and complete axes. Reference-only pages 12–15 were text-read but not visually inspected. No external videos, model card, code, or supplemental materials were inspected.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 16, 17, 18. 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; I. Introduction; II. Related Work (PDF pp. 1–3)
- III. Preliminaries (p. 3)
- IV. RECAP, A–C (pp. 3–5)
- V. Implementation, Model, and System Details, A–D; Algorithm 1 (pp. 5–7)
- VI. Experimental Evaluation, A–C (pp. 7–11)
- VII. Discussion and Future Work; Acknowledgements (p. 11)
- References (pp. 11–16)
- Appendix A. Contributions; B. Additional Value Function Visualization; C. Computing the log-likelihood for policy improvement (pp. 16–17)
- Appendix D. PPO implementation; E. Using CFG for test-time policy improvement; F. Additional algorithm details (pp. 17–18)
Outside the original text pass
- Identity: the title page verifies arXiv:2511.14759v2, dated 19 November 2025, with the supplied title and complete author list. The catalog submission date is 18 November 2025. Version 1 was not supplied, so revision differences beyond the observed version/date cannot be established; mathematical title typography is normalized in the supplied observedTitle.
- 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.
- All eight supplied text chunks were read individually, covering all 18 PDF pages. Original figures were inspected separately in the PDF; the extraction limitation above was addressed by that visual pass.
- The linked project, accompanying videos, model card, code, checkpoints, and training datasets were not supplied or inspected. No experiments were reproduced. Reference-only pages 12–15 were read as text, not rendered for visual review.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e01PDF p. 1, title/author block, arXiv margin stamp, Abstract and Figure 1
Title and Physical Intelligence author roster match the supplied catalog; the margin identifies 2511.14759v2, 19 November 2025. The abstract introduces RECAP and experience-based VLA improvement.
Go to primary source ↓e02PDF pp. 2–3, Sections I–IV; p. 3, Preliminaries and footnote 1
The method combines demonstrations, autonomous trials, and interventions. Preliminaries define returns and n-step advantages and explicitly simplify observations to Markovian states.
Go to primary source ↓e03PDF p. 4, Section IV-A, Equation (1), and preceding pre-training paragraph
A multi-task distributional critic learns discretized empirical returns using cross-entropy with 201 bins. Expected bin values yield a continuous critic of the dataset behavior mixture. Pre-training uses tens of thousands of demonstration hours.
Go to primary source ↓e04PDF p. 5, Section IV-B, Equations (2)–(3), intervention paragraph and footnote 2
Positive advantage is defined by A>ε_ℓ; the objective fits conditional and unconditional behavior. Corrective actions are forced positive. β=1 gives direct conditional sampling; high guidance can cause aggressive behavior.
Go to primary source ↓e05PDF p. 4, Figure 3 and caption; p. 6, Sections V-A–B
The figure separates VLA and value model, and its advantage arrow enters the VLA through binarization. The VLA uses Gemma 3 4B and an 860M action expert with stop-gradient insulation. Subtask text precedes I, FAST actions, and continuous actions; the expert does not consume FAST predictions.
Go to primary source ↓e06PDF p. 7, Section V-C, Equation (5), Figure 5 and caption
Rewards are zero for terminal success, −C_fail for terminal failure, and −1 otherwise. Values are task-length normalized; the value backbone is 670M. The evaluation robot has two 6-DoF arms, parallel-jaw grippers, three cameras, and 50-Hz joint-position control.
Go to primary source ↓e07PDF p. 6, Algorithm 1; p. 7, Section V-D
SFT fixes the indicator positive. Experience is aggregated, then value and policy are fine-tuned from pre-trained checkpoints rather than the previous iteration. Corrections address major mistakes but are not guaranteed optimal.
Go to primary source ↓e08PDF p. 18, Appendix F, Advantage Estimation, Advantage conditioning dropout, Advantage threshold
Post-training uses N=50; pre-training uses a full-episode empirical-return expression. Indicator dropout is 30%. Threshold targets are about 30% positives for pre-training, generally 40% for fine-tuning, and 10% for T-shirts/shorts.
Go to primary source ↓e09PDF p. 8, Section VI-A and Figure 6
Task definitions specify 200-second simple-laundry, strict-laundry and double-espresso limits, 500-second button-up-shirt evaluation despite 11-type training, and a 600-second box-assembly limit.
Go to primary source ↓e10PDF p. 9, Figure 7, Section VI-B final-model paragraph and VI-C.1
Throughput counts successful tasks/hour, with standard-error bars. Approximate final versus SFT bars are 60 versus 33 (simple laundry), 8.3 versus 3.8 (diverse), 29 versus 14.5 (espresso), and 13 versus 10 (boxes). Default evaluation uses β=1.
Go to primary source ↓e11PDF p. 9, Figure 8, Section VI-C metric definitions and results
Success labels aggregate human quality judgments; bars show standard errors. Espresso is approximately 92% versus 39% for SFT. Box success is broken into four task stages. The plotted stages do not all improve monotonically despite the caption generalization.
Go to primary source ↓e12PDF p. 10, Figures 9–10 and Section VI-C.2
The text describes 300 simple-laundry trajectories on four robots per iteration without corrections. Figure 9 has approximate i=0,1,2 throughputs of 33,42,60 for laundry and 10,8.5,13 for boxes; prose states 50% and 2× improvements without a matching explicit endpoint comparison.
Go to primary source ↓e13PDF p. 9, Section VI-B AWR/PPO definitions; p. 10, Figure 11 and VI-C.3; p. 17, Appendix D
Same RECAP-collected data are used for extraction comparisons. Figure 11 shows approximately 60,30,22 successes/hour for RECAP, AWR, PPO. Baselines start from π0.6 without advantage conditioning; PPO uses a single-step diffusion likelihood bound and an SPO-style trust-region variant.
Go to primary source ↓e14PDF p. 10, Figure 12; p. 11, Section VI-C.4
The strict collar-up experiment reports 97% success after two iterations, with 600 trajectories per iteration and an explicit no-intervention/no-additional-demonstration claim. Figure 12 shows an initial SFT success bar near 23%.
Go to primary source ↓e15PDF p. 11, Section VII, Discussion and Future Work
Authors identify human reward labeling, interventions and resets, largely greedy exploration, and batch offline updates rather than a fully concurrent online loop as limitations.
Go to primary source ↓e16PDF p. 6, Equation (4) and Section V-B; p. 16, Appendix C, Equations (6)–(9)
Continuous likelihood is not closed form; flow-matching loss motivates a lower bound alongside token likelihood. Appendix Equation (6) writes I in the subtask factor, while Section V-B says only action likelihoods are affected because I follows the subtask.
Go to primary source ↓e17PDF pp. 17–18, Appendix E, Equations (12)–(13) and continuation
Conditional/unconditional guidance approximates sharpened policy sampling. High β can induce overly aggressive motions; moderate β∈[1.5,2.5] is suggested where useful.
Go to primary source ↓e18PDF p. 18, Appendix F, Dataset composition
Task-dependent collection includes diverse laundry: 450 evaluation and 287 correction episodes; cafe: 414 autonomous and 429 correction episodes in one iteration. Failure and box counts conflict with the main-text accounts.
Go to primary source ↓e19PDF p. 5, Figure 4 and caption; pp. 16–17, Appendix B and Figure 13
Value traces show local decreases during mistakes and increases during recovery/progress. These are illustrated trajectories rather than quantitative critic calibration or held-out prediction tests.
Go to primary source ↓e20PDF p. 7, Section V-D threshold paragraph; p. 18, Appendix F, Advantage threshold
Main text selects the 30th percentile of predicted values; appendix instead selects a threshold giving about 30% positive demonstration advantages on a random 10k-point sample. No reconciliation is provided.
Go to primary source ↓e21PDF p. 10, Section VI-C.2; p. 11, Section VI-C.4; p. 18, Appendix F, Dataset composition
Main failure-removal account says 600 trajectories per iteration with no corrections; appendix says about 1000 autonomous plus 280+378 correction episodes over three robots. Main box account calls 600 trials autonomous; appendix calls 600 episodes demonstrations, both with 360 corrections per iteration.
Go to primary source ↓8.5 Primary sources
π∗ 0.6: a VLA That Learns From Experience ↗
PDF · 13,569 extracted words
Source fingerprint
9f4f5a3070ad9549ea7687ea54baa19858820dbcc0834cc805212d68fd6c8cb3