FlowDAgger: Human-in-the-Loop Adaptation of Generative Robot Policies in Latent Space
1. Paper overview
In one sentence: FlowDAgger turns expert actions into supervision for a frozen robot policy's noise input, improving adaptation while remaining limited by the behaviors that generator can express. overviewinversionadaptationjoint-inversionmetaworldhardwareretentionedm
| At a glance | What to know |
|---|---|
| Research problem | Source description Deployment exposes manipulation failures outside pretraining coverage. Full-policy fine-tuning is costly and can impair existing skills; autonomous latent-space reinforcement learning needs exploration and rewards. FlowDAgger asks whether corrective actions can directly supervise a frozen policy's noise input on states encountered during deployment. overviewgenerative-map |
| Core mechanism | Source description Per-step action inversion supplies latent targets from ordinary expert actions, allowing interactive imitation through an existing generative policy's noise interface. inversionadaptation |
| A key reported result | MetaWorld adaptation across twelve tasks: 0.78; +0.25 over base 0.53. Mean success rate. pi0.5; 50 additional rollouts; three seeds, 25 evaluation rollouts each. SFT uses demonstrations; DAgger methods use corrections; DSRL uses autonomous reward. SFT 0.71; LoRA-DAgger 0.68; Residual-DAgger 0.64; DSRL 0.55. Best reported mean and eight task wins; a 25-percentage-point gain. Matching episodes does not equalize human information or labor. protocolmetaworld |
| Reading caution | Source description The frozen generator bounds representable corrections. Sparse or biased interventions limit coverage; ill-conditioned or multimodal dynamics can impair inversion. limits |
Core contributions
- Source description
Per-step action inversion supplies latent targets from ordinary expert actions, allowing interactive imitation through an existing generative policy's noise interface. inversionadaptation
- Source description
The extension to Cosmos-Policy inverts joint action, future-state and value latents. The experiments span action-head VLAs, a world-action model and a diffusion policy. joint-inversionfamiliesgr00tdiffusion
Figure 1. Correct the action, recover its noise, then update the small policy that chooses future noise. Original paper, p. 2 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Follow the upper path from observation s into the latent policy and then the frozen generator. The generator produces an action a; an expert can supply a corrective action a* to the environment. Follow that correction down into Action Inversion, which finds w* whose decoded action approximates a*. The lower arrow sends w* left into latent-space supervised training, and the dashed update arrow returns only to the latent policy. The generator has no weight-update arrow. Section 4.3 additionally specifies equal sampling from correction targets and successful autonomous-rollout noises; that second training buffer is omitted from this overview diagram. overviewadaptationedm
What it supports. The useful separation is between changing the generator and changing its input. Human actions become labels for an observation-conditioned noise policy, while the base still produces the actions. This makes the intervention actionable without requiring the operator to understand latent coordinates or provide a reward function.
Where the evidence stops. The figure writes w sampled from the latent policy and draws a distribution; Section 4.3 and Equation (4) specify deterministic deployment. This guide follows the equation. The schematic also omits Cosmos's terminal backpropagated inversion solve.
2. Motivation
2.1 The problem and the proposed response
Deployment exposes manipulation failures outside pretraining coverage. Full-policy fine-tuning is costly and can impair existing skills; autonomous latent-space reinforcement learning needs exploration and rewards. FlowDAgger asks whether corrective actions can directly supervise a frozen policy's noise input on states encountered during deployment. overviewgenerative-map
2.2 What this reading follows
A human can correct a robot's motion, but cannot directly specify the noise that a generative policy should receive. FlowDAgger bridges that gap by reversing the frozen generator and using the recovered noise as a training target. A small observation-conditioned policy then chooses the noise for future actions. Read the mechanism first, then separate three experimental questions: whether adaptation improves, whether existing skills survive, and whether inversion quality matters. The reported gains span simulation and physical manipulation, but the retention table shows residual forgetting. The world-action extension also needs a terminal gradient-based solve that qualifies the introduction's simpler account. overviewinversionadaptationjoint-inversionmetaworldhardwareretentionedm
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 | Foundational work |
| Architecture | Not applicable |
| Prediction paradigm | Not applicable |
| Quadrant | Not applicable |
This table preserves the labels recorded at reading time. The current major category is WAMs. View the current classification.
3.1 Evidence-based assessment
Supports the recorded classification
The recorded foundational-method classification fits a training/adaptation technique applied to several base architectures. FlowDAgger itself is neither a single new world-action backbone nor inverse dynamics: it reverses a generator numerically to obtain noise labels. Cosmos's shared denoiser jointly predicts world/action quantities, but that property cannot classify the entire framework as One Model. Architecture, prediction paradigm and quadrant being not applicable is supported. generative-mapinversionjoint-inversionadaptationfamilies
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 an expert takeover into a reusable noise decision
The operator corrects behavior in the robot's action space. FlowDAgger first pairs that action with the observation at the intervention, then asks which noise would make the frozen generator reproduce it. This is a label-construction problem: the expert never needs to choose a latent vector. Once inversion produces w*, the observation and target enter an accumulated intervention buffer. A separate buffer stores the generating noises from successful autonomous rollouts. Equal proportions of the two buffers train an observation encoder and MLP with squared-error regression. Deployment then composes that small policy with the unchanged generator. Later interventions occur on states reached by the updated policy, so the training distribution changes with its behavior. The source's deterministic deployment rule is more specific than the sampling symbol shown in Figure 1. overviewinversionadaptation
5.2 Reverse the discrete step that actually produced the action
For an action-head model, the numerical sampler advances a latent by adding a velocity times the step size. Reversing the process requires the velocity at the unknown preceding latent. A single explicit reverse step instead evaluates it at the known next latent. FlowDAgger solves the resulting implicit equation by repeatedly updating a candidate for the preceding latent, then moves to the next reverse step. Section 4.1 gives a sufficient contraction condition, delta-t L below one; it does not establish that condition globally for every correction. The default five inner iterations are empirically motivated by Table 6: increasing to ten only reduces MSE from 0.00168 to 0.00147 while time rises from 456 to 614 ms. Table 5 then asks whether cleaner targets improve downstream execution. generative-mapinversioniterationsablation
5.3 For a world-action model, invert the joint generator
Cosmos-Policy cannot be treated as an isolated action decoder: one denoiser generates action, future-state and value frames in a shared latent tensor. FlowDAgger starts with the base model's clean prediction, inserts the expert-minus-base delta into the action frame, and leaves the other target frames unchanged. It then reverses the entire joint process. The EDM schedule ends at a nonzero noise level, so reversing the terminal denoise requires a short Adam solve through one denoiser call at a time. This differentiates through the frozen model without updating its weights. Because the recovered joint noise is large, the default small policy predicts sixty-four PCA coefficients fitted from warmup targets. Table 2 tests resulting action success; it does not establish the accuracy or consistency of the retained future-state predictions. joint-inversionedmwam-parameterizationfamilies
5.4 Training and inference
During training
A separate observation encoder and MLP regress inverted noise targets. Corrections aggregate across deployment rounds. Each batch mixes intervention targets and generating noises from successful autonomous rollouts equally; the latter provide rehearsal on already successful behavior. The base parameters never update. adaptation
Action-head noise is regressed directly. Cosmos defaults to predicting 64 PCA coefficients using a basis fitted on inverted targets from a brief base-policy warmup. Inversion still operates on full joint noise; the basis's stability advantage over full regression is qualitative. wam-parameterization
During inference
For each observation, the deterministic noise policy supplies the frozen generator's input, which produces the action to execute. New observations close the interaction loop. Inversion supplies training labels when corrections occur; it is not a separate deployment planner. Figure 1 uses sampling notation, but Equation (4) explicitly specifies deterministic noise selection. overviewadaptation
5.5 Implementation flow
- Undo action generation
Start from the corrective action and reverse each Euler step by fixed-point iteration. Repeatedly reevaluate the velocity at the candidate preceding latent rather than using one explicit reverse pass. The default is five inner iterations; the sufficient convergence condition is conditional, not a measured global guarantee. generative-mapinversioniterations
- Invert Cosmos jointly
Construct a target from the base clean latent, adding the expert-minus-base action delta only to its action frame. Keep state/value targets at the base predictions, but invert the full 16×9×28×28 tensor through the shared denoiser. This uses predicted world states without requiring expert future-state labels. joint-inversion
- Handle the terminal denoise
Cosmos's EDM schedule runs from noise level 80 to 4 before a terminal denoise. A local Adam solve reverses that terminal transform, backpropagating through one denoiser call per iteration while keeping weights frozen. Thus the introduction's forward-passes-only statement does not cover the full Cosmos implementation. edm
6. Experiments & results
FlowDAgger converts expert corrections into noise targets for a small policy that steers a frozen generative robot controller. It supports action-head models and joint world-action diffusion. Experiments report efficient adaptation and better retention than the tested alternatives, with meaningful residual forgetting and limited implementation disclosure.
6.1 Read the original evidence
Table 1. Noise-space imitation leads the aggregate comparison, with task-specific exceptions. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start with the bottom two rows: Mean SR averages task success rates, and the delta row subtracts the frozen base. Then scan individual tasks to see whether an aggregate advantage is universal. Bold identifies the best non-base method within each task; ties can occur. Section 5.1 supplies the protocol missing from the crop: fifty additional rollouts, three seeds, and twenty-five evaluation rollouts per seed. Those additional rollouts contain different supervision: offline demonstrations for SFT, corrections for the DAgger methods, and autonomous experience with reward for DSRL. The correction-based baselines are the closest comparisons for the choice of adaptation space. protocolmetaworldfamilies
What it supports. FlowDAgger reports 0.78 mean success against 0.53 for the base, a 25-percentage-point increase, and wins eight of twelve tasks. Its 0.78 exceeds SFT's 0.71 and DSRL's 0.55. Door Lock illustrates a counterexample: Residual-DAgger reaches 0.85 while FlowDAgger reaches 0.75.
Where the evidence stops. The table omits uncertainty estimates. Equal episode budgets do not imply equal human labor or information. The seven-task means in Table 2 use another subset and must not be substituted for these twelve-task means.
Table 4. Real robot gains include Toolbox Packing's increase from 0.13 to 0.80 with ten additional episodes. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read each row as a distinct hardware task with its own adaptation budget. The parenthesized delta compares FlowDAgger to the base; it does not compare against SFT. The final column counts additional episodes, and the caption states evaluation over thirty rollouts. Section 5.1 describes FR3 Duo and Dual UR5e platforms and both single-arm and bimanual tasks, but does not assign every row to a particular base model and platform. These are executed manipulation outcomes rather than generated-video quality scores. Keep the episode unit explicit: the prose's phrase 'ten corrections' for Toolbox does not reveal how many interventions occurred inside those episodes. hardwareprotocol
What it supports. Toolbox Packing reports 0.80 success for FlowDAgger, 0.63 for SFT and 0.13 for the base. Every row improves over the base. Compared with SFT, however, Button Push ties at 0.73 and Jenga Stacking is lower at 0.86 versus 0.90, so the table supports a qualified hardware advantage.
Where the evidence stops. Plug Insertion's reported 0.72 is not a simple integer success count divided by thirty; the source does not explain aggregation. No confidence intervals or per-task base assignments are given. Reported values are preserved without inferring missing counts.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| MetaWorld adaptation across twelve tasks pi0.5; 50 additional rollouts; three seeds, 25 evaluation rollouts each. SFT uses demonstrations; DAgger methods use corrections; DSRL uses autonomous reward. | 0.78; +0.25 over base 0.53. Mean success rate | SFT 0.71; LoRA-DAgger 0.68; Residual-DAgger 0.64; DSRL 0.55. Best reported mean and eight task wins; a 25-percentage-point gain. Matching episodes does not equalize human information or labor. protocolmetaworld |
| Cosmos-Policy adaptation across seven shared MetaWorld tasks Table 2's seven-task subset; default Section 5.1 protocol. | 0.74 versus base 0.53; +0.21. Mean success rate | pi0.5 improves from 0.53 to 0.79 on this subset. Supports compatibility with joint world-action generation; does not measure future-state prediction accuracy. protocolfamiliesjoint-inversion |
| Held-out skill retention after Hammer adaptation pi0.5 adapted on Hammer for 50 episodes; five held-out tasks; SFT uses 50 demonstrations and matched gradient-step budget. | Held-out 0.88 versus base 0.96; Hammer 0.84 versus 0.40. Held-out mean success; Hammer success | Held-out means: Residual-DAgger 0.69, LoRA-DAgger 0.30, SFT 0.02. Best tested retention, with an eight-point mean loss. Push declines from 0.96 to 0.56, so frozen weights do not guarantee preserved behavior. protocolretention |
| Real-world Toolbox Packing Table 4; ten additional episodes; caption specifies 30 evaluation rollouts. Per-task base-model/platform assignment is not supplied. | 0.80 versus base 0.13; reported +0.67. Task success rate | SFT 0.63. Physical manipulation evidence. Episodes must not be equated with individual corrections; uncertainty and exact aggregation are unspecified. hardware |
| Action-inversion procedure comparison Same 29 scripted Assembly chunks; pi0.5, K=10; warmed-up RTX 5090; downstream three-seed, 25-rollout evaluations. | Per-step M=5: 0.00168 MSE, 456 ms; Assembly/Hammer 0.87/0.96. Action MSE; median milliseconds per chunk; downstream success | Euler: 0.0329, 315 ms, 0.59/0.71. Trajectory FP: 0.0228, 1662 ms. Adam: 0.0275, 1350 ms. More accurate labels accompany better execution. Per-step inversion is slower than Euler but faster than the two stronger alternatives; this ablation is a separate protocol from Table 1. ablation |
| Gr00t N1.7 adaptation on LIBERO-90 task 57 LIBERO-10 checkpoint; held-out cream-cheese-to-tray task; 50 episodes; one seed, 15 rollouts per point. | Reaches 1.00 at episode 20 from base 0.60, with later declines. Success-rate trajectory | DSRL fluctuates near 0.70. Suggestive transfer across action-head VLAs; the curve does not establish sustained ceiling performance or seed robustness. gr00t |
| Diffusion Policy adaptation on robomimic LIFT State-only DDPM policy; 50 episodes; one seed, 15 rollouts per point. | Caption reports approximately 0.97 by episode 25 versus approximately 0.48 base. Success-rate trajectory | DSRL levels near 0.80. A second generative family shows gains; values are approximate caption/curve summaries, not pooled multi-seed estimates. diffusion |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Table 3. Freezing the base reduces forgetting in this comparison; it does not eliminate it. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read Hammer separately from the five held-out columns. Hammer is the task used for adaptation, while Door, Drawer, Faucet, Plate and Push probe effects elsewhere. The Mean and delta columns summarize only the held-out tasks. Compare each adapted row with the unadapted base, then inspect individual tasks rather than stopping at the mean. The experiment uses fifty Hammer episodes; SFT receives fifty demonstrations with a matched gradient-step budget. The original caption says bold held-out values are within base noise or above, but does not define that tolerance. The crop preserves the source's typography without converting it into a significance claim. retentionadaptationprotocol
What it supports. FlowDAgger improves Hammer from 0.40 to 0.84 while retaining a held-out mean of 0.88, compared with 0.96 for the base. That exceeds the other adapted means of 0.69, 0.30 and 0.02. Nevertheless, Push falls from 0.96 to 0.56, revealing a substantial task-specific regression beneath the strong average.
Where the evidence stops. This comparison bundles the frozen generator, learned noise policy and successful-rollout rehearsal. It does not isolate which component preserves skills. A single adaptation task and five held-out tasks also leave broader retention untested.
Table 5. Inversion accuracy matters downstream, while the cheapest reverse pass remains faster. Original paper, p. 12 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Move across each row from reconstruction quality to computation and then executed task success. Lower Action MSE and time are better; higher SR is better. The caption fixes twenty-nine scripted-expert Assembly chunks, four-dimensional environment actions with each dimension ranging from minus one to one, and warmed-up median timings on an RTX 5090. Appendix B.1 fixes the base decoder's ten-step Euler schedule. The final two columns evaluate policies trained on the respective inverter targets using three seeds and twenty-five rollouts. This separation matters: reconstructing a correction accurately is an intermediate diagnostic; succeeding after supervised adaptation is the control outcome. ablationmetaworld
What it supports. Per-step fixed point with five inner iterations reports 0.00168 MSE, versus 0.0329 for Euler reverse, and higher downstream success on both listed tasks. It takes 456 ms per chunk: slower than Euler's 315 ms but faster than Adam's 1350 ms and trajectory fixed point's 1662 ms.
Where the evidence stops. The downstream protocol is separate from Table 1, so its success rates are not interchangeable. The caption identifies Assembly reconstruction chunks but leaves Hammer target collection underspecified. Improved reconstruction and success are linked here without establishing a universal error threshold.
7. Analysis & limitations
7.1 What the evidence leaves open
The frozen generator bounds representable corrections. Sparse or biased interventions limit coverage; ill-conditioned or multimodal dynamics can impair inversion. limits
Table 4 ties SFT on Button Push and loses on Jenga, qualifying the surrounding claim of outperforming SFT except on Jenga. Plug Insertion's 0.72 cannot be reconstructed as a simple success count out of 30; aggregation remains unresolved. hardware
No confidence intervals accompany the main tables. Figure 3 shading is seed min/max. Neither reduced forgetting nor staying on a behavioral manifold is a universal guarantee from these comparisons. metaworldretentionlearning-curveslimits
7.2 Questions for discussion
- How much retention comes from frozen weights versus successful-rollout rehearsal?
- When a corrected action changes the future, how consistent are Cosmos's retained base state/value targets?
- Would the advantage persist under matched human intervention time rather than matched episodes?
8. Reproducibility audit
8.1 Requirements and known gaps
A minimal reconstruction check needs the frozen pi0.5 decoder, its ten-step schedule, expert chunks and action normalization. Preserve separate inverter timing and downstream evaluation. Table 6 lists 60 default model evaluations, whereas Section 4.1 counts KM=50 velocity calls; the difference is not explicitly reconciled. inversionablationiterations
Reported pi0.5 adaptation memory is approximately 8 GB; the inversion microbenchmark uses an RTX 5090. Neither establishes Cosmos memory or full control-loop latency. Exact noise-encoder architecture, regression optimizer, update schedule, checkpoint identifiers for the main bases, warmup size and terminal-Adam settings remain insufficiently specified. computeablationadaptationprotocoledmwam-parameterization
Proposed controls should compare inversion procedures on identical corrections and independently remove successful-rollout rehearsal while measuring Hammer improvement and every held-out task. These would test label fidelity and retention mechanisms separately. ablationadaptationretention
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 cleaner inversion supervision cause better execution?
Reader-proposed, not run: freeze one pi0.5 checkpoint and its ten-step sampler, and invert the same held-out set of scripted Assembly corrections with Euler reverse, per-step M=5 and M=10, and trajectory k=5. Log action MSE, total decoder evaluations and warmed-up runtime separately. Train otherwise identical noise policies with equal correction sets, rehearsal buffers and update budgets, then evaluate Assembly and Hammer on fixed initial-state sets across multiple seeds. Test the Table 6 matched-evaluation pair M=10 versus k=5 explicitly. If the reconstruction advantage persists but downstream success does not improve, the claimed control benefit of cleaner labels is weaker than the microbenchmark suggests. generative-mapadaptationablationiterations
Check 2: Separate frozen-weight retention from successful-rollout rehearsal
Reader-proposed, not run: repeat Hammer adaptation with the same base, inverted correction targets and optimizer budget. Compare the published equal-mixture buffer against correction-only training, keeping the base frozen in both arms. Use fixed prerecorded correction data for the first comparison so changed intervention opportunities cannot explain differences; follow with matched-budget on-policy collection. Evaluate Hammer and every Table 3 held-out task separately, especially Push, using the same seeds and initial states. Similar Hammer gains but worse held-out retention without rehearsal would support the buffer's proposed role. Indistinguishable retention would weaken that explanation; persistent Push regression would bound the preservation claim. adaptationretentionprotocol
8.3 Reading coverage
Visual audit: Actually viewed the title/author/version page (1), Figure 1 and introduction (2), generative equations (3), inversion and joint latent definitions (4), joint target, training/inference and setup (5), task photographs, learning curves and protocol (6), Tables 1–2 and Figure 4 (7), Tables 3–4 and limitations (8), EDM and PCA implementation details (11), inversion comparison (12), iteration sweep and Gr00t curve/setup (13), and diffusion-policy curve (14). All six original figures and all six tables were visually inspected. Five final crops were separately viewed; the narrow hardware table was rendered at higher DPI and recropped to remove caption fragments. Figure 1's sampling notation was checked against deterministic Equation (4); the guide discloses the discrepancy. Reference pages 9–10 were read in the complete text chunks but not rendered. Separate supplements and code remain outside this reading.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 11, 12, 13, 14. 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 page and Abstract
- 1 Introduction
- 2 Related Work
- 3 Background
- 4 Method, including 4.1–4.3
- 5 Experiments, including 5.1–5.5
- 6 Limitations
- 7 Conclusion
- References
- Appendix A.1–A.3: World-Action Model Inversion Details
- Appendix B.1–B.4: Additional Experiments
Outside the original text pass
- 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.
- The image-extraction limitation was addressed by inspecting PDF pages 1–8 and 11–14, all six figures, all six tables, and every final crop. Reference pages 9–10 were read as text.
- Only arXiv:2607.08877v1, dated 9 July 2026, was supplied. Its title and ten authors match the catalog; no other edition was available for comparison. Appendix A.2 describes an earlier inverter implementation, not an established separate paper revision.
- Code, linked websites and separate supplements were not inspected; no experiments were reproduced.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
identityPDF p. 1, title, author/affiliation block and arXiv margin
The title matches the catalog. The ten authors, in order, are Michael Murray, Daphne Chen, Simran Bagaria, Dean Fortier, Tess Hellebrekers, Galen Mullins, Harshavardhan Gajarla, Oier Mees, Maya Cakmak, and Andrey Kolobov. Affiliations are Microsoft Research, Microsoft, ETH Zurich, and University of Washington. The margin identifies arXiv:2607.08877v1 [cs.RO], 9 July 2026.
Go to primary source ↓overviewPDF pp. 1–2, Abstract, Section 1 and Figure 1/caption
FlowDAgger maps corrective actions into the noise space of a frozen generative robot policy and trains a separate latent policy. Figure 1 routes expert actions through inversion and noise targets through supervised training; its update arrow returns to the latent policy. The diagram shows sampling notation, whereas Section 4.3 specifies deterministic deployment.
Go to primary source ↓generative-mapPDF p. 3, Section 3, Equations (1)–(2)
The observation s is the full policy input, possibly including sensor history; a is an action chunk of horizon H. A learned velocity field maps initial noise w to an action through an ODE, discretized into K Euler steps with step size 1/K.
Go to primary source ↓inversionPDF p. 4, Section 4.1, Equation (3)
Each forward Euler step is inverted by fixed-point iteration starting from the next latent. The stated sufficient contraction condition is delta-t times L less than one. The default uses M=5, with K=10 in the discussed action-head regime. This inverse uses KM velocity evaluations and no base-model backpropagation.
Go to primary source ↓joint-inversionPDF pp. 4–5, Section 4.2
Cosmos-Policy jointly generates action, future-state and value frames in a 16×9×28×28 latent tensor with a shared denoiser. The inversion target starts from the base clean latent and adds the expert-minus-base action delta only to the action frame; the state/value target frames remain at the base predictions. Inversion nevertheless recovers full joint noise.
Go to primary source ↓adaptationPDF p. 5, Section 4.3, Equations (4)–(5)
The only trained component is a separate observation encoder plus MLP noise policy. Deployment uses its deterministic output as input to the frozen base. Corrections are inverted and aggregated across rounds; regression batches contain equal shares of intervention targets and generating noises from successful autonomous rollouts.
Go to primary source ↓protocolPDF pp. 5–6, Section 5.1 and Baselines and protocol
Simulation uses MetaWorld; real experiments use FR3 Duo and Dual UR5e platforms. The default adaptation budget is 50 additional rollouts, with three seeds and 25 evaluation rollouts per seed. SFT receives offline demonstrations, correction-based methods receive human-gated corrections, and DSRL receives autonomous experience and sparse reward.
Go to primary source ↓metaworldPDF p. 7, Table 1, task rows, Mean SR and delta rows; p. 6, Section 5.2
Across twelve MetaWorld tasks with pi0.5, reported mean success is 0.53 base, 0.71 SFT, 0.68 LoRA-DAgger, 0.64 Residual-DAgger, 0.55 DSRL and 0.78 FlowDAgger. FlowDAgger wins eight rows. Door Lock favors Residual-DAgger, 0.85 versus 0.75; Assembly is 0.89 versus base 0.64.
Go to primary source ↓familiesPDF p. 7, Table 2 and Section 5.3
On seven shared MetaWorld tasks, pi0.5 improves from 0.53 to 0.79 mean success, while Cosmos-Policy improves from 0.53 to 0.74. Their reported gains are 0.26 and 0.21 respectively. These means use a different task subset from Table 1.
Go to primary source ↓retentionPDF p. 7, Section 5.4; p. 8, Table 3, Hammer and held-out columns
Adaptation is on Hammer for 50 episodes; SFT uses 50 demonstrations with matched gradient-step budget. FlowDAgger raises Hammer success from 0.40 to 0.84 and has held-out mean 0.88 versus base 0.96. Residual-DAgger, LoRA-DAgger and SFT have held-out means 0.69, 0.30 and 0.02. FlowDAgger's Push score drops from 0.96 to 0.56. The caption bolds held-out scores described as within base noise or above, without defining that uncertainty threshold.
Go to primary source ↓hardwarePDF p. 8, Section 5.5 and Table 4, caption and all task rows; p. 5, Section 5.1
Table 4 reports real-hardware success over 30 rollouts and five to twenty additional episodes. Toolbox Packing is 0.13 base, 0.63 SFT and 0.80 FlowDAgger with ten additional episodes. Button Push ties SFT at 0.73; Jenga Stacking is 0.86 versus SFT 0.90. Plug Insertion is reported as 0.72 with twenty episodes. Section 5.5 also calls Toolbox's ten episodes ten corrections, without specifying their within-episode count.
Go to primary source ↓limitsPDF p. 8, Section 6
The authors limit adaptation by the frozen policy's representable behavior, intervention quality and state coverage, and inversion reliability under multimodal or poorly conditioned generative dynamics. They acknowledge that additional data or weight adaptation may be needed.
Go to primary source ↓edmPDF p. 11, Appendix A.1; p. 2, Section 1
Cosmos uses an EDM schedule from sigma_max=80 to sigma_min=4 followed by a terminal denoise. Fixed-point reversal handles Euler integration, but a local Adam solve recovers the terminal denoiser input and backpropagates through one denoiser call per step. Base weights remain frozen. This is an explicit exception to the introduction's forward-passes-only description.
Go to primary source ↓wam-parameterizationPDF p. 11, Appendices A.2–A.3
The earlier action-frame-only optimizer has reported reconstruction floors of about 0.04 on base actions and 0.076 on expert actions, with insufficient metric/protocol detail for comparison to Table 5. Current inversion acts on full joint noise. The default Cosmos noise policy predicts 64 PCA coefficients from a basis fitted to inverted-noise warmup targets. Full regression also worked; the claimed stability advantage of the basis is qualitative.
Go to primary source ↓ablationPDF p. 12, Table 5/caption and Appendix B.1
Inversion comparison uses the same 29 scripted-expert Assembly chunks with pi0.5 and a K=10 Euler sampler. Action MSE is measured in four-dimensional environment actions ranging from -1 to 1 per dimension. Median warmed-up RTX 5090 times are 315 ms Euler reverse, 1350 ms Adam, 1662 ms trajectory fixed point and 456 ms per-step fixed point. Corresponding MSEs are 0.0329, 0.0275, 0.0228 and 0.00168. Per-step downstream Assembly/Hammer success is 0.87/0.96, versus Euler 0.59/0.71; evaluation uses three seeds and 25 rollouts.
Go to primary source ↓iterationsPDF pp. 12–13, Appendix B.2 and Table 6
Per-step M=3,5,10 gives MSE 0.00397, 0.00168, 0.00147 and time 405,456,614 ms, respectively. Table 6 lists 40,60,110 model evaluations. At 110 listed evaluations, trajectory k=5 gives MSE 0.0228 and 1662 ms. The table's default 60 evaluations is not explicitly reconciled with Section 4.1's KM velocity-call count.
Go to primary source ↓gr00tPDF p. 13, Appendix B.3 and Figure 5/caption
Gr00t N1.7 uses the authors' LIBERO-10 checkpoint and is adapted on held-out LIBERO-90 task 57, cream cheese into tray. The comparison allocates 50 episodes, scripted interventions for FlowDAgger and reward for DSRL. The source describes a ten-component DCT-II projection of 40×132 chunk noise. Figure 5 is single-seed, 15 rollouts per point, with a 60% base. FlowDAgger reaches the ceiling at episode 20 but later declines; DSRL fluctuates near 70%.
Go to primary source ↓diffusionPDF pp. 13–14, Appendix B.4 and Figure 6/caption
The additional DDPM-based Diffusion Policy experiment uses robomimic LIFT with low-dimensional state observations, 50 adaptation episodes, one seed and 15 evaluation rollouts per point. The caption reports approximately 48% base and approximately 97% FlowDAgger by episode 25; DSRL levels near 80%.
Go to primary source ↓computePDF p. 6, Compute efficiency; p. 7, Figure 4/caption
The pi0.5 adaptation footprint is reported as approximately 8 GB, similar to deploying that base model. Figure 4 compares Assembly success at 50 episodes against peak training VRAM. This is not a Cosmos memory or end-to-end latency benchmark.
Go to primary source ↓learning-curvesPDF p. 6, Figure 3 and caption
Five MetaWorld learning curves compare FlowDAgger, Residual-DAgger and DSRL against dashed frozen-base rates. Curves average three seeds with 25 rollouts per evaluation; shaded regions denote per-seed minima and maxima, not confidence intervals.
Go to primary source ↓8.5 Primary sources
FlowDAgger: Human-in-the-Loop Adaptation of Generative Robot Policies in Latent Space ↗
PDF · 6,899 extracted words
Source fingerprint
0f7fc3c1792ef4645ef866526736a83a5e2cf4f55be513158bb2e165c62df1dd