VAMPO: Policy Optimization for Improving Visual Dynamics in Video Action Models
1. Paper overview
In one sentence: Rewarding completed video latents can improve the early predictive features used by a separate action policy, but most measured control gains require retraining that policy. e02e03e05e06e09
| At a glance | What to know |
|---|---|
| Research problem | Author claim The authors argue that likelihood-surrogate video training can produce plausible futures with small pose, spatial-relation or contact-timing errors that mislead manipulation. VAMPO targets this mismatch between prediction and control. Its reward still measures latent consistency with demonstrations, rather than executed task return. e02e06 |
| Core mechanism | |
| A key reported result | L-CALVIN ABC→D long-horizon manipulation: 6.73; 0.39 Average completed trajectory length; ten-task completion fraction. Ten-task chains under ABC→D. VPP: 5.53; 0.28. Gains are 1.20 tasks and 0.11 absolute completion fraction. These ten-task results are separate from CALVIN percentages. Trial counts and uncertainty are not reported. e08e11 |
| Reading caution |
Core contributions
Figure 2. Post-training evaluates completed futures while concentrating the stochastic policy update at the first transition. Original paper, p. 5 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start with the observation and instruction on the left, then follow the upper blue predictor from noisy latent to completed latent. The SDE region marks the first stochastic transition; subsequent ODE steps complete each rollout deterministically. On the lower right, a visual encoder maps the expert future to the target latent. The reward blocks compare it with the completed candidates, and the upward arrow sends the optimization signal back to the predictor. Equation 16 is essential here: maximize negative L1 error plus cosine similarity, despite the positive L1 block drawn in this schematic. The separate action model is described in Sections 2.1–2.2, not drawn here. e03e05e06e07
What it supports. This is a training diagram: expert future observations supply reward targets. At deployment, the AGM uses first-step hidden features and language. The figure therefore illustrates how terminal prediction quality is intended to improve an earlier representation, rather than an online robot planner choosing among decoded future videos.
Where the evidence stops. The positive mean-L1 block and plus sign conflict with Equation 16’s negative L1 norm; reduction conventions remain unclear. Fire/snowflake symbols also do not document separate frozen networks. Equations 15 and 18 establish deterministic completion and a first-transition policy ratio.
2. Motivation
2.1 The problem and the proposed response
The authors argue that likelihood-surrogate video training can produce plausible futures with small pose, spatial-relation or contact-timing errors that mislead manipulation. VAMPO targets this mismatch between prediction and control. Its reward still measures latent consistency with demonstrations, rather than executed task return. e02e06
2.2 What this reading follows
A robot can fail even when its predicted future looks plausible: a small spatial error can make a grasp arrive at the wrong place. VAMPO asks whether video predictors should be post-trained for agreement with expert futures before their features are passed to a robot policy. Its distinctive choice is to randomize only the first denoising transition, then finish each candidate deterministically and compare its latent with the demonstration. This reading follows the reward back to the early features, separates prediction evidence from executed manipulation, and examines the missing controls behind the claimed connection between better visual dynamics and better actions. e02e03e05e06e09
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
Dual-system × IDM fits the explicit VPM-to-AGM architecture: future features condition a separate inverse-dynamics action generator. VAMPO post-trains prediction rather than jointly emitting future observations and actions from one model. Its reward operates during training, not inference-time search. e03e05e06
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 Separate the future target from the feature used for action
VAMPO contains two prediction problems. The VPM uses the current observation and language to model a future video latent. During training, a VAE encoding of the demonstrated future provides the target. The action model consumes a different object: multilayer hidden states from the VPM’s first denoising step, together with language. It learns to reconstruct demonstrated actions through its own diffusion process. This distinction explains why a sharper final video is not automatically a better controller: the action model never needs to wait for that whole video to be decoded. The authors therefore place the stochastic policy decision at the early transition, then evaluate its eventual consequences after deterministic completion. The downstream link still has to be checked through actual action execution, as Table 1 attempts to do. e03e05e06e09
5.2 Follow one group through the reward update
Algorithm 1 samples conditions and expert future latents, then shares an initial noisy latent across the candidate group. Each candidate receives its own first-transition stochastic draw, followed by deterministic completion. Equation 16 scores the completed latent with negative L1 error and positive cosine similarity; Table S2 sets both weights to one and the group size to eight. Equation 17 centers and scales those rewards within the group, so the update favors candidates that outperform their peers for the same condition. The clipped GRPO objective then uses the first transition’s new-to-old policy likelihood ratio. As a reader interpretation, this restricts where stochastic credit is assigned, but it does not prove that terminal latent reward is sufficient supervision for every control-relevant early feature. Figure 2’s positive L1 drawing must not override the equation. e05e06e07e17
5.3 Test each link in the proposed explanation
The experiments offer three different kinds of evidence. Figure 3 evaluates prediction alignment and selected decoded futures. Table 1 evaluates executed CALVIN task chains and a Jacobian-based coupling diagnostic. Figure 5 evaluates partial-credit physical manipulation, while Table 3 extends simulated chain length. As reader analysis, these observations are consistent with the proposed path from improved prediction to useful features to improved action. They do not uniquely identify it: the final model also receives downstream AGM training, and the table lacks a matched AGM-only control. The ablations favor first-step stochasticity and latent rewards, but provide no direct reward-hacking measurement or uncertainty. A reproduction should therefore preserve these separate outcome categories and evaluate transfer explicitly instead of treating a rising reward or effective rank as a proxy for successful execution. e09e11e12e15e16e19e22
5.4 Training and inference
During training
Simulation starts from VPP. CALVIN ABC contains 18,033 trajectories; a selected subset produces 129,454 clips. VPM post-training uses about 1,500 steps on 64 H20 GPUs, followed by roughly 10 AGM epochs on 8 H20 GPUs. Appendix A specifies ABC for AGM training; the main text says the whole CALVIN dataset. e08e17
Table S2 specifies group size 8, batch size 8, clipping 0.2, fine-tuning/post-training learning rates 10⁻⁴/10⁻⁶, and both reward weights 1.0. Table S1 lists 16-frame prediction, 256×256 images, a 10×7 action shape and 10 diffusion-transformer sampling steps. e17
During inference
Extract multilayer VPM hidden states at the first denoising step and condition the separate DDIM-style action head on these features and language. Expert futures and GRPO scoring belong to training. Action-chunk execution frequency and replanning schedule are unspecified. e03e06e17
5.5 Implementation flow
- Predict in latent space
A Stable Video Diffusion VPM conditions on observation and instruction. A pretrained VAE encodes the expert future clip as the clean target. Initial supervised training reconstructs clean video latents from noisy inputs. e03
- Define denoising actions
The MDP state contains the noisy latent, noise level and condition. Its action is the next latent, not a robot command. Reward is terminal, after the future latent has been completed. e04
- Branch once, complete deterministically
Candidates share an initial noisy latent. Euler-Ancestral noise is injected at the first denoising transition; remaining transitions use Euler ODE updates. The Gaussian first transition supplies the likelihood ratio for policy optimization. e05e06
- Optimize and transfer
Compare completed candidates with the expert latent using negative L1 error plus cosine similarity. Normalize rewards within the group and maximize a clipped GRPO objective. Subsequently train the diffusion AGM on predictive features and demonstrated actions. e03e06e17
6. Experiments & results
VAMPO post-trains a video predictor with rewards for agreement with expert future latents, then trains a separate action policy on its predictive features. First-step stochastic denoising concentrates optimization where the action policy obtains its representation. The supplied v1 reports stronger simulated manipulation and higher real-robot task scores, but latent agreement does not directly measure contact accuracy or guarantee control success.
6.1 Read the original evidence
Figure 3. Latent alignment improves overall, accompanied by selected examples of corrected predictions. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the left panel as a latent prediction diagnostic: training steps run horizontally and L1 loss vertically. The vertical scale shows a narrow range, so a large visual drop should not be mistaken for a large absolute change. The curve fluctuates early, declines, and then varies around a lower level. On the right, gray, green and red borders identify ground truth, VAMPO and the baseline. Compare corresponding image positions inside each of the three columns, rather than treating neighboring examples as one continuous execution. The labels identify the authors’ interpretations of those selected prediction differences. e08e12
What it supports. The curve and examples support better agreement with expert future representations in this experiment. They make the proposed intermediate mechanism plausible: the predictor can suppress an extraneous object or correct a spatial trajectory. They do not by themselves show that a robot completed any of these tasks.
Where the evidence stops. The loss plot reaches 2,000 steps, while the training recipe says about 1,500; their run correspondence is unspecified. No uncertainty or evaluation-set construction accompanies the curve, and the examples are selected qualitative evidence rather than measured pose/contact errors.
Table 1. An unchanged action head benefits modestly; subsequent action-head training produces a larger gain. Original paper, p. 9 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read downward to separate representation replacement from downstream adaptation. The first row is the base policy; the second changes the VPM while preserving the original AGM; the third also trains the AGM. Columns 1–5 report completing progressively longer task chains, followed by average completed length. These are CALVIN ABC→D results, so the evaluation environment differs from the training environments. The two rightmost columns describe the effective rank of the action-versus-vision Jacobian and its normalized ratio. Dashes in the middle row are missing measurements, not zero coupling. Use the completion columns to assess control and the rank columns as a separate sensitivity diagnostic. e08e09
What it supports. With the original AGM, average length increases from 4.28 to 4.35 and five-task completion from 74.7% to 77.2%. Training the AGM after VPM post-training raises them to 4.56 and 83.1%. The larger final gain thus combines representation improvement with downstream adaptation.
Where the evidence stops. There is no original-VPM plus newly trained-AGM row, so additional action training is not isolated. Higher effective rank is not proof of useful causal dependence, and no repeated-seed uncertainty is supplied for these comparisons.
Table 3. The longest evaluated chains show a larger absolute gain in average completed length. Original paper, p. 10 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Begin with the Train→Test column: every row belongs to the ABC→D protocol. Then scan from one completed task toward ten; the entries are completion fractions, unlike the percentage formatting in the five-task CALVIN table. The far-right column summarizes average completed trajectory length over the longer horizon. VPP is the most informative base comparison because VAMPO starts from that framework. The highlighted row remains above VPP throughout the chain lengths shown, while both rows decline as more consecutive tasks are required. Keep the ten-task threshold and the average-length statistic separate: they answer different questions about how far execution progresses. e08e11
What it supports. VAMPO reports average length 6.73 versus VPP’s 5.53, an absolute gain of 1.20 tasks. Ten-task completion rises from 0.28 to 0.39, or 11 percentage points. This supports better long-chain performance in the reported held-out environment without implying that most ten-task chains are completed.
Where the evidence stops. The table does not report evaluation counts, seed variation or confidence intervals. It establishes results for L-CALVIN’s protocol; it cannot show robustness to arbitrary longer sequences or equal training compute across all listed model families.
Figure 5. The physical robot results concern partial-credit manipulation scores on three tasks. Original paper, p. 11 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. The left panel identifies the Agibot platform and camera placement; the upper-right images show the apple, bowl and two-bottle tasks. Read each bar chart only against its own task: the light bar is baseline VPP and the dark bar is VAMPO. The vertical label is Score, not success rate. Appendix B assigns 50 points each for grasp and placement in the first two tasks, and 25 points for each arm’s grasp and placement in the bottle task. Thus partial completion earns credit. The figure prints no exact bar values, and the report preserves that precision limit rather than converting bar heights into exact measurements. e13e15e16
What it supports. VAMPO’s score bar is higher for every displayed task, providing physical execution evidence beyond the paper’s generated-video comparisons. The bottle task remains visibly more difficult under this scoring. These results support an improvement within the demonstrated setup, while the missing evaluation repetitions prevent a statistical assessment.
Where the evidence stops. Section 4.2 specifies D405 wrist cameras, but this figure labels them RealSense 455. Appendix B describes bowl placement onto a plate despite the box label. Evaluation trial counts, score aggregation and uncertainty are unspecified.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| CALVIN ABC→D manipulation Train on ABC; evaluate five-task chains in unseen D. | 4.56; 83.1% Average completed trajectory length; five-task completion rate | VPP: 4.28; 74.7%. Updated VPM with original AGM: 4.35; 77.2%. Tri-VLA: 4.37; 81.8%. The VPP comparison gains 0.28 tasks and 8.4 percentage points. Most improvement accompanies subsequent AGM training; there is no AGM-only retraining control. Cross-family results do not establish matched compute. e08e09e10 |
| L-CALVIN ABC→D long-horizon manipulation Ten-task chains under ABC→D. | 6.73; 0.39 Average completed trajectory length; ten-task completion fraction | VPP: 5.53; 0.28. Gains are 1.20 tasks and 0.11 absolute completion fraction. These ten-task results are separate from CALVIN percentages. Trial counts and uncertainty are not reported. e08e11 |
| Vision–action coupling diagnostic CALVIN base versus post-trained VPM/AGM; Jacobian of actions with respect to visual features. | 43.88; 0.0814 Average effective rank; average normalized effective-rank ratio | Base: 29.28; 0.0603. The paper uses squared sum of singular values divided by sum of their squares. Increased rank suggests broader visual sensitivity; it does not establish useful causal dependence. e09 |
| CALVIN sampler, reward and optimizer ablations ABC→D, comparing each reported variant with the final configuration. | One-step SDE, latent reward and GRPO: 4.56 Average completed trajectory length | Five-step SDE: 4.34; pixel reward: 4.39; DDPO: 4.36; base: 4.28. Selected components outperform alternatives in these runs. Table S3 prints +0.08 for 4.39 versus 4.28, although the displayed difference is 0.11. Reward hacking remains an explanation, not a separately measured outcome. e18e19e21e22 |
| Agibot Genie 01 tabletop manipulation Apple grasp-and-place, bowl placement and dual-arm bottle placement; 200 demonstrations collected per task. | Figure 5 shows higher VAMPO scores on all three tasks; exact scores are not printed. Task score with partial credit; maximum 100 per Appendix B | Base VPP bars are lower in every panel. These are execution scores, not binary success percentages. Trial counts, error bars and aggregation details are absent. Bowl labels say box, whereas Appendix B describes placement onto a plate. e13e15e16 |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Figure 6. Component comparisons favor the final design, with training instability visible only in the appendix table. Original paper, p. 12 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Each panel changes one reported design choice against the base policy and the final configuration. The first five bar groups use the left axis for task-chain success percentages. The last group, separated by a dashed line, uses the right axis for average length; its bar height cannot be compared directly with the success bars. The top-left panel shows only 400 and 1,400 post-training steps. Consult Table S5 for the omitted intermediate checkpoints, where performance dips. The lower-left panel tests the location/count of stochastic transitions, while the lower-right compares the space in which prediction consistency is rewarded. e18e19e20e21e22
What it supports. The final configuration reaches 4.56 average length versus 4.34 for five-step SDE, 4.39 for pixel reward and 4.36 for DDPO. These comparisons support the selected sampler, reward space and optimizer in the reported CALVIN runs. They do not directly measure whether later stochastic steps exploit the reward.
Where the evidence stops. No error bars are supplied. Table S3’s printed pixel improvement is +0.08 although 4.39 minus 4.28 is 0.11. Table S5’s 600-step average falls to 4.31; the selected checkpoints in this figure omit that instability.
7. Analysis & limitations
7.1 What the evidence leaves open
Latent agreement and decoded examples do not directly measure pose or contact accuracy. Figure 3 shows nonmonotonic loss; Figure S2 acknowledges imperfect texture and detail. The largest control gain changes both predictor and action-model training. e06e12e23e09
Table S5 gives average lengths 4.50, 4.31, 4.42 and 4.56 at 400, 600, 1,000 and 1,400 steps. Average length always exceeds the base here, but some 600-step short-chain rates fall below it. Repeated-seed uncertainty and checkpoint selection rules are absent. e20e22
Video comparisons span Agibot, Flexiv and a platform labeled VidowX; executed-task evidence is supplied for Agibot. Figure 5 labels wrist cameras RealSense 455, while Section 4.2 specifies D405. These are unresolved source discrepancies. e13e14e15
7.2 Questions for discussion
- Does matched AGM-only retraining explain part of the gain attributed to predictive features?
- Can latent reward improve while first-step features become less useful for control?
- How sensitive are results to the chosen expert future when several futures are valid?
8. Reproducibility audit
8.1 Requirements and known gaps
Reproduction requires VPP initialization, the exact clip subset and preprocessing, reward reduction, noise schedule, old-policy refresh interval and gradient-detachment implementation. The PDF supplies architecture/hyperparameters but leaves these details incomplete; its code link alone establishes no reproducibility. e01e05e06e08e17
Proposed checks should pair first-step versus five-step stochasticity at matched budgets, and cross original/updated VPMs with original/retrained AGMs. Log latent error, task completion and uncertainty to separate reward improvement, transfer and extra action training. These are unexecuted proposals. e06e09e19
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 first-step sampler improve control at a matched budget?
Reader-proposed experiment, not performed: initialize identical VPP copies and compare one-step versus five-step SDE post-training on the same CALVIN ABC clip subset, with shared evaluation seeds, G=8, identical reward weights, update counts and separately recorded denoiser-call budgets. Use Equation 16’s reward sign and document its tensor reduction. Evaluate each VPM first with the same frozen AGM, then with AGMs trained for the same budget. Log terminal reward, held-out latent L1/cosine, CALVIN average length and five-task completion with uncertainty across seeds. The credit-assignment account predicts that one-step sampling produces better action transfer even when terminal rewards are comparable. If its advantage disappears under matched budgets, or reward rises without control gains, the claimed explanation needs qualification. e05e06e07e08e09e17e19
Check 2: Separate representation improvement from extra action training
Reader-proposed factorial check, not performed: compare original versus VAMPO-post-trained VPM features, each paired with an original AGM and an AGM retrained on the same demonstrations for the same number of updates. Keep preprocessing, task sequences and action execution settings fixed. This supplies the missing original-VPM/retrained-AGM control. Measure the five-task and ten-task completion curves, average length, and the same action-versus-vision Jacobian diagnostic; record the exact effective-rank normalization. Test whether the updated VPM adds improvement beyond retraining alone. If an AGM-only control matches the full system, or rank grows without improved execution, the report’s proposed causal chain is weakened. Replicate the comparison across seeds before interpreting small frozen-AGM gains. e03e08e09e11e17
8.3 Reading coverage
Visual audit: Visually inspected the title/author page, all method and equation pages, Figures 1–6, Tables 1–3, Appendix A and Tables S1–S2, Appendix B scoring and Figure S1, Tables S3–S6, and Figure S2. All six final original-PDF crops were individually viewed with labels, legends and relevant table columns intact. Figure 2 was checked against Eqs. 15–19 and Algorithm 1; reward-sign and reduction discrepancies are disclosed. Figure 5 camera and task labels were checked against pp. 10 and 18. Appendix numerical values were checked against Figure 6. All five text chunks, including reference pages 14–16, were read; those reference-only pages were not visually inspected. No external supplement, code, linked model or experiment was inspected or executed.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 17, 18, 19, 20. 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 and affiliation block; arXiv:2603.19370v1, 19 March 2026 (p. 1)
- Abstract and Section 1 Introduction (pp. 1–3)
- Section 2.1–2.3: paradigms, instantiation and objective mismatch (pp. 3–4)
- Section 3.1–3.3: MDP, SDE sampling, hybrid sampler, rewards and Algorithm 1 (pp. 5–7)
- Section 4.1–4.3: simulation, real-world evaluation and component analysis (pp. 7–11)
- Section 5 Related Work and Section 6 Conclusion (pp. 11–13)
- References (pp. 14–16)
- Appendix A Training Details (p. 17)
- Appendix B Real-world Experiment Details (pp. 17–18)
- Appendix C Ablations (pp. 18–19)
- Appendix D Visualization (p. 20)
Outside the original text pass
- Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
- Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout. This extraction limitation was addressed by inspecting the original PDF figures and tables.
- Separate supplemental material availability has not been fully verified.
- Only the supplied v1 was reviewed; no other revision or edition was supplied or compared. Its title and all 19 authors agree with the catalog.
- Code, linked models and external resources were not inspected; no experiments were reproduced.
- All five supplied text chunks were read individually. PDF pages 1–13 and 17–20 were visually inspected; reference-only pages 14–16 were read as text.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e01PDF p. 1, title/author/affiliation block and arXiv margin
Title matches VAMPO: Policy Optimization for Improving Visual Dynamics in Video Action Models. The 19 authors match the catalog in order. The margin identifies arXiv:2603.19370v1 [cs.RO], 19 Mar 2026. Thirteen affiliations and code/model/project links are printed.
Go to primary source ↓e02PDF pp. 2 and 4, Section 1 and Section 2.3
Authors motivate post-training by a mismatch between likelihood surrogates and precise pose, contact and spatial-relation information needed for manipulation.
Go to primary source ↓e03PDF pp. 3–4, Sections 2.1–2.2, Eqs. (1)–(6)
Separate SVD-based VPM and diffusion AGM; expert video is VAE-encoded; first-step multilayer hidden states condition the AGM with language. Video and action denoisers use reconstruction objectives; the action head uses DDIM-style denoising.
Go to primary source ↓e04PDF p. 5, Section 3.1, Eqs. (7)–(9)
MDP state includes latent, noise level and condition; action is the next latent; terminal reward evaluates the completed output.
Go to primary source ↓e05PDF p. 6, Sections 3.2–3.3, Eqs. (10)–(15)
Euler-Ancestral sampling supplies a Gaussian transition. Hybrid sampling uses it only at the first transition and deterministic Euler updates thereafter, motivated by first-step feature relevance and credit assignment.
Go to primary source ↓e06PDF p. 7, Algorithm 1 and Eqs. (16)–(19)
Candidates share initial noise, complete hybrid rollouts, receive negative L1 plus cosine rewards against expert latents, and use group-normalized advantages. The clipped likelihood ratio indexes the first transition; old policy refresh is periodic without a stated interval.
Go to primary source ↓e07PDF p. 5, Figure 2 and caption; p. 7, Eq. (16)
Figure 2 shows conditioning, first-step SDE, ODE completion and latent comparison. Its positive mean L1 formula and plus sign omit the negative distance term in Eq. (16); mean versus norm reduction is also not reconciled.
Go to primary source ↓e08PDF p. 8, Section 4.1, Simulation Settings and Training Details
ABC→D uses an unseen environment; L-CALVIN extends chains from five to ten tasks. CALVIN ABC has 18,033 trajectories; a selected subset yields 129,454 clips. VPM uses about 1.5k steps on 64 H20 GPUs and AGM about ten epochs on eight H20 GPUs; evaluation uses RTX 5880.
Go to primary source ↓e09PDF p. 8, Improvement in Action Execution and coupling paragraph; p. 9, effective-rank definition and Table 1
Base / updated VPM with original AGM / updated VPM and AGM have lengths 4.28/4.35/4.56 and five-task rates 74.7/77.2/83.1. ER is (sum of singular values) squared divided by sum of squares. Base/full ER is 29.28/43.88 and ERR 0.0603/0.0814; middle-row coupling entries are missing.
Go to primary source ↓e10PDF p. 9, Table 2, VPP, Tri-VLA and VAMPO rows; comparison paragraph
VAMPO reports 98.0/94.8/91.3/88.3/83.1 and average 4.56; VPP 96.0/91.3/86.4/80.4/74.7 and 4.28; Tri-VLA 96.8/92.4/86.8/83.2/81.8 and 4.37. Authors state no extra network components or data are introduced.
Go to primary source ↓e11PDF p. 10, Table 3, VPP and VAMPO rows, ten-task and Avg. Len columns
L-CALVIN reports VPP ten-task fraction 0.28 and average 5.53, versus VAMPO 0.39 and 6.73. No repeated-run uncertainty is included.
Go to primary source ↓e12PDF p. 8, Figure 3 and Analysis of Visual Dynamics
Latent L1 loss decreases overall with fluctuations; qualitative panels compare expert, VAMPO and baseline under hallucination suppression, planning correction and action rectification labels. These are prediction comparisons.
Go to primary source ↓e13PDF p. 10, Section 4.2, Real-World Settings and Training Details
Execution platform is Agibot Genie 01; 14 arm DoF and 2-DoF gripper per end-effector are described. Three synchronized RGB-D views include D455 head and D405 wrist cameras. Each task has 200 demonstrations. Stages use 1.5k and 10k additional steps without a precise module mapping.
Go to primary source ↓e14PDF p. 11, Figure 4 and caption; p. 10, Section 4.2
Video panels identify Agibot, Flexiv and VidowX. L1 before/after values are respectively 0.2124/0.2014, 0.2694/0.2603 and 0.2304/0.2294. The body separates multi-embodiment video evaluation from Agibot execution.
Go to primary source ↓e15PDF p. 11, Figure 5 and caption
Setup, task examples and baseline/VAMPO score bars cover apple, bowl and two-bottle tasks. VAMPO bars are higher; exact labels, uncertainty and trial counts are absent. Head and wrist labels both say RealSense 455, conflicting with p. 10.
Go to primary source ↓e16PDF p. 18, Appendix B, tasks (1)–(3) and Figure S1
Apple and bowl tasks allocate 50 points to grasp and 50 to placement; bottles allocate 25 to each of two grasps and two placements. Maximum is 100. Bowl naming refers to a box, but the text describes a plate. Evaluation repetition and aggregation are unspecified.
Go to primary source ↓e17PDF p. 17, Appendix A, Tables S1 and S2
ABC is specified for both stages. S1 lists 16-frame video, 10×7 actions, 256×256 images, Video Former and diffusion-transformer configurations, and ten sampling steps for the latter. S2 lists G=8, batch=8, clipping=0.2, fine-tuning/post-training rates 10⁻⁴/10⁻⁶ and unit reward weights.
Go to primary source ↓e18PDF p. 18, Appendix C, Table S3, Avg. Len column
Base, pixel and latent rewards yield 4.28, 4.39 and 4.56. Printed pixel delta +0.08 conflicts with the displayed difference +0.11.
Go to primary source ↓e19PDF p. 19, Table S4; p. 11, Hybrid Denoising Type
Five-step SDE yields 4.34 and 76.4% five-task completion; first-step-only SDE 4.56 and 83.1%; base 4.28 and 74.7%. Reward hacking is the authors’ interpretation.
Go to primary source ↓e20PDF p. 19, Table S5; pp. 10–11, Post-Training Steps
Average lengths at 400/600/1000/1400 steps are 4.50/4.31/4.42/4.56. At 600 steps, one-/two-/three-task rates 95.6/90.6/85.6 are below base 96.0/91.3/86.4, despite higher average length.
Go to primary source ↓e21PDF p. 19, Table S6; p. 11, Optimization Algorithm
DDPO reaches average 4.36 and five-task rate 77.6; GRPO reaches 4.56 and 83.1. Authors hypothesize more stable updates but provide no variance measurement.
Go to primary source ↓e22PDF p. 12, Figure 6 and caption
Four panels compare 400/1400 steps, DDPO/GRPO, five/one SDE steps and pixel/latent reward. Success-rate and average-length bars use different vertical axes. Intermediate checkpoints appear only in Table S5.
Go to primary source ↓e23PDF p. 20, Appendix D, Figure S2 and caption
Two one-step prediction sequences compare ground truth, VAMPO and baseline. The caption acknowledges imperfect textures and detail while asserting improved motion dynamics.
Go to primary source ↓8.5 Primary sources
VAMPO: Policy Optimization for Improving Visual Dynamics in Video Action Models ↗
PDF · 8,086 extracted words
Source fingerprint
23b9ec1da2991b51e5ddfe5f03f36531b24a9170e16c6dadb17b9a989cb10dff