RoboScape: Physics-informed Embodied World Model
1. Paper overview
In one sentence: Depth feedback and tracked-token supervision improve action-conditioned robotic video prediction, but their benefits trade across metrics and do not yet establish physical robot transfer. e02e04e05e08e09e10e11e13
| At a glance | What to know |
|---|---|
| Research problem | Source description Robotic demonstrations are costly, while visually plausible generated frames can still deform objects or break motion continuity. The paper asks whether depth and tracked-point supervision can improve embodied video prediction without an inference-time physics-engine cascade. e02e04 |
| Core mechanism | |
| A key reported result | Pi0 policy training on LIBERO: 66.6%, 71.4%, 74.4%, 79.1% with 200, 400, 600, 800 synthetic examples Average task success rate. 200 real trajectories provide warm-up before increasing generated data; Table 3 averages Spatial, Object, Goal, and LIBERO-10. Real (200): 65.2%; the largest synthetic condition gains 13.9 percentage points. This supports augmentation after real-data warm-up, not synthetic-only learning. Training-data quantity differs between conditions. e09 |
| Reading caution |
Core contributions
- Source description
An automated data pipeline adds estimated depth and keypoint tracks, filters clips, and groups them by action difficulty and scene for curriculum learning. e03
Figure 2. Geometry enters through depth-to-RGB feedback; tracked points provide an additional training constraint. Original paper, p. 4 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start with the two lower-left image streams. Snowflakes mark the visual tokenizers as frozen, and the action embedding splits into both prediction branches. Follow the dashed gray arrows upward: Equation 3 confirms that depth features are projected into RGB features, rather than the reverse. The current-time frames on the right supply token targets for training alignment. Above them, the adaptive sampler chooses moving tracks whose predicted features are compared across time. Figure 2 and Equation 5 both use the predicted initial token as the reference. Temporal causal attention and spatial bidirectional attention are specified in the accompanying text, not drawn as explicit masks. e04e05e06
What it supports. The physical prior affects the predictor through two distinct routes: learned depth features participate in RGB prediction, while tracked-point consistency shapes training. The red regularization arrow is not an action-output connection. A separately supplied robot action conditions the visual future in both branches.
Where the evidence stops. The diagram does not specify how initial depth is obtained at inference or fully document the tokenizer training setup. Frozen-tokenizer icons and a unified diagram are insufficient to reconstruct those implementation choices.
2. Motivation
2.1 The problem and the proposed response
Robotic demonstrations are costly, while visually plausible generated frames can still deform objects or break motion continuity. The paper asks whether depth and tracked-point supervision can improve embodied video prediction without an inference-time physics-engine cascade. e02e04
2.2 What this reading follows
Imagine asking a video model what a robot will see after its gripper moves. A convincing next image is useful only if objects also move and deform consistently. RoboScape adds two kinds of supervision to that prediction problem: depth supplies scene geometry, while tracked points encourage persistent features on moving objects. The resulting model still receives actions from outside; separate policies use its generated data or interact with its predicted observations. Read the evidence in three layers: frame prediction, policy training, and policy evaluation. Each tests a different capability, and the ablation and scaling results reveal why none alone certifies physical understanding. e02e04e05e08e09e10e11e13
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 | Not assigned |
| Architecture | Not assigned |
| Prediction paradigm | Not assigned |
| Quadrant | Not assigned |
This table preserves the labels recorded at reading time. The current major category is Benchmarks & simulators. View the current classification.
3.1 Evidence-based assessment
Insufficient evidence to decide
The catalog snapshot is unassigned, so no recorded quadrant can be confirmed. Architecture evidence supports an action-conditioned forward video world model with coupled RGB/depth branches. Actions are supplied externally; joint RGB-depth training does not establish a single model jointly predicting future observations and executable actions, or an inverse-dynamics action decoder. e04e05e10
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 Follow an action into a predicted observation
Begin with the paper’s forward-dynamics formulation: the next RGB observation depends on observation history and supplied continuous robot actions. MAGVIT-2 compresses RGB and depth into token grids, so the Transformer predicts compact representations rather than operating directly on full-resolution images. Actions and positions are added to both branches, keeping the same control signal available to RGB and depth prediction. Temporal attention is causal, while spatial attention is bidirectional within a frame. The depth branch then injects projected intermediate features into RGB at successive blocks. During training, current frames provide targets; during the reported rollout, the model starts from one conditioning frame and predicts fifteen more. This world model predicts consequences. The downstream policy remains the component that chooses actions. e04e06e10
5.2 Why follow points instead of freezing pixels?
A moving object changes image coordinates even when it remains the same object. RoboScape therefore compares features along estimated point trajectories, using the initial predicted feature as the reference for later frames. The adaptive sampler favors tracks with large accumulated displacement, concentrating supervision on active regions. A second objective increases RGB token weights along those trajectories. Reader interpretation: the first constraint encourages persistence through movement, while the second directs learning effort toward hard moving regions; neither measures material stiffness or enforces a conservation law. This distinction matters when reading the ablation. Removing keypoints weakens appearance and action sensitivity but improves depth metrics. The data-scaling appendix presents a similar metric conflict, motivating a check that separates motion modeling from static-image accuracy. e05e08e13
5.3 Separate useful training data from a reliable evaluator
The downstream studies ask two different questions. In policy training, generated observations contribute data from which a separate policy learns: Diffusion Policy uses synthetic-only Lift data, whereas pi0 first receives real LIBERO trajectories for warm-up. The success rates test whether those learning inputs help. In policy evaluation, an already trained policy instead interacts with the world model, whose generated observations stand in for the environment. Success judged from those videos is compared with simulator success across saved checkpoints. Reader analysis: high correlation can support ranking checkpoints even if the model systematically underestimates strong policies, as the scatter plot suggests. Neither downstream study demonstrates physical robot transfer, and manual video scoring introduces a separate uncertainty from the world model’s prediction error. e09e10e11
5.4 Training and inference
During training
Section 3.1 reports 50,000 AgiBotWorld-Beta videos/clips spanning 147 tasks and 72 skills, then approximately 6.5 million 16-frame training clips sampled at 2 Hz. It reports five epochs, lambda1=1, lambda2=0.01, lambda3=1, gamma=5, and about 24 hours on 32 NVIDIA A800-SXM4-80GB GPUs. e06
During inference
Condition on the first frame and supplied actions, then autoregressively predict 15 subsequent frames. The depth branch provides internal geometric feedback. In policy evaluation, a separately trained policy supplies actions and receives generated observations; success is judged manually from model rollouts. e04e06e10
5.5 Implementation flow
- Construct supervision
Video Depth Anything estimates depth; SpatialTracker supplies tracks. TransNetV2 detects camera boundaries; Intern-VL supplies action semantics and keyframes; FlowNet helps filter motion. These are preprocessing dependencies despite the unified prediction model. e03
- Encode observations and actions
MAGVIT-2 tokenizes RGB and depth into spatial latent grids; Figure 2 marks the visual tokenizers as frozen. Actions concatenate end position, end orientation, and effector position. Learned action and positional embeddings are added to each branch’s historical tokens. e04e06
- Predict with geometric feedback
Both branches use temporal causal attention and spatial bidirectional attention. Each depth-block feature is linearly projected and added to the corresponding RGB feature. RGB and depth token targets receive cross-entropy supervision. e04
- Emphasize moving regions
Rank densely tracked points by summed inter-frame displacement and keep the top K. Align their predicted token features to the initial-frame predicted features. A separate RGB loss weights trajectory-intersected locations by gamma and other locations by one; this is loss weighting, not an inference-time controller. e05
6. Experiments & results
RoboScape predicts action-conditioned robotic RGB and depth videos using coupled autoregressive branches. Depth-feature feedback and motion-selected keypoint supervision aim to improve geometry and interaction dynamics. Reported benefits include stronger video metrics, useful synthetic policy-training data, and correlated simulator-based policy evaluation; the ablations expose tradeoffs, and physical robot deployment remains future work.
6.1 Read the original evidence
Table 1. RoboScape leads the listed baseline entries across appearance, depth, and action sensitivity. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the metric directions before comparing rows. Lower LPIPS measures better perceptual similarity and lower AbsRel means less relative depth error; PSNR, the two depth accuracy columns, and delta-PSNR favor larger values. The last column measures sensitivity to action conditioning, so it answers a different question from image resemblance. Compare RoboScape with Genie across all three column groups rather than selecting one favorable metric. Keep CogVideoX’s dash as missing information. Section 3.1 describes one conditioning frame and fifteen generated frames; the PDF does not provide the held-out split or evaluation count behind this table. e06e07e12
What it supports. RoboScape reports LPIPS 0.1259 and PSNR 21.8533, compared with Genie’s 0.1683 and 19.7571. Its action-sensitivity score is 3.3435 versus 1.9871. These results support stronger reported video prediction under the authors’ evaluation, while leaving the causal contribution of each auxiliary task to the ablation.
Where the evidence stops. No uncertainty intervals or complete matched-training protocols are supplied. Appendix B identifies Genie as an open-source reproduction, and CogVideoX is text-conditioned; the comparison does not establish equal action interfaces or compute.
Table 3. Synthetic-only Lift training and real-warm-up LIBERO augmentation use different protocols. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Treat the left and right halves as separate experiments. For Robomimic Lift, the body specifies Diffusion Policy trained for 10,000 steps using only generated data. Its final synthetic row can be compared with the 200-real-trajectory row directly as reported. On the right, read the four LIBERO suite columns before the Average column. Section 3.3 adds a condition not printed in the column heading: pi0 receives 200 real trajectories for warm-up before generated data are used. Consequently, the right-hand synthetic rows are evidence for augmentation after real-data training, and should not be described as synthetic-only policy learning. e09e11
What it supports. At 200 synthetic examples, DP reaches 91% versus 92% with 200 real trajectories. For pi0, the largest generated-data condition reaches 79.1% average success versus 65.2% for the real-data comparator, a 13.9-percentage-point difference. Both experiments show increasing reported success as generated data increase.
Where the evidence stops. The LIBERO comparison changes total training-data quantity, and the table supplies no uncertainty intervals or policy-evaluation trial counts. It does not isolate synthetic data quality from data volume or demonstrate physical robot deployment.
Figure 5. RoboScape tracks checkpoint quality better than the listed alternatives, without matching absolute success rates exactly. Original paper, p. 9 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Each dot is a saved Diffusion Policy checkpoint, not a separate manipulation task. Read simulator success horizontally and success judged from generated video vertically; the dashed diagonal marks equality. Labels identify training epochs, while the final label marks convergence. The experiment uses Robomimic Lift policies trained on 200 trajectories, checkpoints every 250 epochs, and 100 evaluation runs in each environment. Success comes from the simulator automatically in one setting and manual video judgments in the other. Compare the rising pattern in RoboScape with the nearly flat low-success predictions of the two baselines, then inspect distance from the diagonal separately. e10
What it supports. The reported Pearson correlation is 0.953 for RoboScape, compared with -0.134 for IRASim and -0.195 for iVideoGPT. Its reported R-squared is 0.908. As reader analysis, the stronger ordering agreement is useful for selecting checkpoints, while the offsets from the diagonal warn that ordering and calibration are different properties.
Where the evidence stops. These points come from one task and one policy-training trajectory. Manual success labeling, unspecified post-training details, and absent confidence intervals limit generalization; the reported R-squared should not be read as proof of probability calibration.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| Action-conditioned robotic video prediction AgiBotWorld-Beta-derived evaluation; first frame conditions 15 predicted frames. Held-out split and evaluation sample count are not specified. | 0.1259 / 21.8533; 0.3600 / 0.6214 / 0.8307; 3.3435 LPIPS / PSNR; AbsRel / delta1 / delta2; delta-PSNR | Genie: 0.1683 / 19.7571; 0.4425 / 0.5435 / 0.7736; 1.9871. CogVideoX has no reported delta-PSNR. Best reported entries across these columns among the listed baselines. Action sensitivity is a proxy, not a robot success rate; uncertainty is unreported. e06e07e12 |
| Physics-supervision ablation Table 2 removes depth, keypoint learning, or both from the video model. | Full: 0.1259 / 21.8533 / 0.3600 / 3.3435 LPIPS / PSNR / AbsRel / delta-PSNR | Without depth: 0.1249 / 21.9465 / 0.3921 / 3.4863. Without keypoints: 0.1264 / 21.7087 / 0.3417 / 2.9462. Depth improves AbsRel relative to its removal, but slightly worsens appearance and action sensitivity. Keypoints improve appearance and action sensitivity, but worsen the three reported depth metrics. This is a tradeoff, not uniform component-wise dominance. e08 |
| Diffusion Policy training on Robomimic Lift DP trained for 10,000 steps using only generated data; real-data comparator uses 200 trajectories. | 40%, 77%, 84%, 91% with 50, 100, 150, 200 synthetic examples Task success rate | 200 real trajectories: 92%. The largest synthetic condition is one percentage point below the real comparator. No error bars establish equivalence. e09 |
| Pi0 policy training on LIBERO 200 real trajectories provide warm-up before increasing generated data; Table 3 averages Spatial, Object, Goal, and LIBERO-10. | 66.6%, 71.4%, 74.4%, 79.1% with 200, 400, 600, 800 synthetic examples Average task success rate | Real (200): 65.2%; the largest synthetic condition gains 13.9 percentage points. This supports augmentation after real-data warm-up, not synthetic-only learning. Training-data quantity differs between conditions. e09 |
| Policy evaluation against a ground-truth simulator Robomimic Lift DP trained on 200 trajectories; checkpoints every 250 epochs; post-trained world model; 100 runs in each evaluation environment. | 0.953; 0.908 Pearson correlation; reported R-squared | IRASim: -0.134; 0.018. iVideoGPT: -0.195; 0.038. World-model success uses manual judgment, simulator success uses simulator signals. Strong checkpoint correlation does not establish calibrated success probabilities or cross-task reliability. e10 |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Table 2. The component study reveals competing objectives rather than uniform gains. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Use the whole-model row as the reference, then move vertically within each column. Removing depth increases AbsRel from 0.3600 to 0.3921, consistent with a geometric benefit from depth supervision. However, the same removal improves LPIPS, PSNR, and delta-PSNR. Removing keypoint learning changes the pattern: appearance and action sensitivity worsen, but AbsRel and both depth-threshold scores improve. The last row removes both designs and should be read separately rather than assumed to equal either single removal. Figure 4 provides a selected qualitative example of local distortions; it does not overturn these aggregate tradeoffs. e08e05
What it supports. The full model’s delta-PSNR is 3.3435 versus 2.9462 without keypoints, supporting an action-sensitivity benefit. Yet the no-keypoint model has better AbsRel, 0.3417 versus 0.3600. The defensible conclusion is complementary but competing effects, not that every auxiliary component improves every reported measure.
Where the evidence stops. The study does not separately isolate tracked-token consistency from trajectory-weighted RGB loss, nor provide seed variability or a parameter-matched control. Small appearance differences cannot be treated as statistically established improvements.
Figure 7. Increasing training data improves appearance and action sensitivity while worsening the plotted depth metrics. Original paper, p. 15 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. The caption and Appendix C describe this as a data-size study of RoboScape-S. Follow each metric across 1000K, 3000K, and 6000K clips, remembering that lower is better for LPIPS and AbsRel. LPIPS improves while PSNR and delta-PSNR rise; the depth panels move in their unfavorable directions. This differs from Figure 6’s model-capacity experiment, which uses 34M, 131M, and 544M variants. The authors suggest that smaller datasets encourage copying the final conditioning frame, which can inflate geometric scores while missing dynamics. That explanation is an author interpretation of the pattern, not a separately controlled result shown here. e13e07
What it supports. The visual makes geometric score improvements an unreliable standalone indicator of better dynamics. In this experiment, more data yields stronger action sensitivity and appearance even as depth accuracy worsens. A meaningful follow-up would examine moving regions and static regions separately before treating any one metric as physical fidelity.
Where the evidence stops. The retained legend lists S/M/L despite the fixed-S caption and text; it does not establish capacity changes here. CogVideoX’s plotted zero action-sensitivity line also differs from Table 1’s missing entry and is not a verified numeric result.
7. Analysis & limitations
7.1 What the evidence leaves open
Real-world robot testing is explicitly future work. Selected frame strips illustrate successes and failures but cannot establish continuous physical validity or deployment reliability. e11e14
Appendix C reports better metrics with larger models (34M, 131M, 544M). Increasing RoboScape-S data from 1M to 3M to 6M clips improves appearance/action sensitivity while degrading plotted depth metrics; the authors attribute this to small-data static-frame overfitting, without a dedicated causal control. e13
Evaluation split construction, uncertainty, precise depth-evaluation procedure and delta-PSNR calculation are absent. Baselines are heterogeneous, and two recent geometric world models are excluded because their training code was unavailable to the authors. e06e07e12
7.2 Questions for discussion
- Would motion-stratified evaluation separate genuine dynamics learning from static-frame shortcuts?
- Would blinded success labeling and new-task checkpoints preserve the policy-evaluation correlation?
8. Reproducibility audit
8.1 Requirements and known gaps
Reconstruction requires the annotation models, AgiBot data processing, both tokenizers, coupled Transformers, and policy-task adaptation. The PDF leaves K, initial dense sample count, filtering thresholds, optimizer, learning rate, batch size, software versions, and precise tokenizer setup unspecified. e03e04e05e06
The conversion from 50,000 source clips to 6.5 million training windows lacks a complete sampling recipe. Depth initialization at inference, synthetic action/trajectory construction, and downstream world-model post-training also need clarification before a faithful reproduction. e04e06e09e10
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: Separate moving-point supervision from static-frame shortcuts
Reader-proposed check: use the same episode-separated data, tokenizer, model size, training budget, action sequences, and seeds for a two-by-two comparison of tracked-token consistency and trajectory-weighted RGB loss, holding the depth branch fixed. Compare top-motion tracks with uniformly sampled tracks at the same K, and include a repeated-conditioning-frame prediction baseline. Report the paper’s metrics separately for moving and static regions plus simulator task success where available. The mechanism is weakened if motion-selected supervision brings no repeatable moving-region or action-response benefit, or if apparent geometric gains are matched by the static baseline. Document chosen missing hyperparameters as reconstruction assumptions. e03e05e06e08e13
Check 2: Test policy ranking and calibration with blinded labels
Reader-proposed check: recreate the Lift checkpoint comparison with 200 training trajectories and 100 runs per checkpoint in each environment, then repeat on an independently trained policy and an additional task. Match initial-state distributions, freeze world-model adaptation before evaluation, and have two raters independently judge anonymized generated videos using a prespecified success rule. Report rater agreement, bootstrapped correlation intervals, checkpoint ranking errors, and absolute success-rate error against simulator signals. Compare RoboScape with the reported baselines. The evaluator claim is weakened if its correlation advantage disappears across policies or raters, even when the original checkpoint sequence remains well ordered. e10e12
8.3 Reading coverage
Visual audit: The title/byline and affiliations, complete method and equations, training hardware/settings, all three quantitative tables, baseline details, scaling appendix, and all twelve original figures were visually inspected. Text reading covered all five supplied chunks, including reference pages 10–13; those bibliography-only pages were not rendered. All six final crops were individually inspected with their headers, axes and legends retained. Figure 2’s depth-to-RGB arrows, frozen-tokenizer icons and predicted first-frame keypoint target agree with the relevant formulation. Figure 7’s legend/caption ambiguity and the missing-versus-zero CogVideoX action-sensitivity entry are disclosed in its caution. The source frame strips were inspected as still images; no continuous video, code or separate supplement was reviewed.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 9, 14, 15, 16, 17. Appendix coverage: reviewed.
Original figures and tables remain the work of the source’s authors. Extractions preserve their scientific content; any HTML wrapper layout is disclosed with each figure. The surrounding reading notes are our own.
Text reading scope & known omissions
- Abstract; 1 Introduction
- 2 Methodology: 2.1 Problem Formulation; 2.2 Robotic Data Processing Pipeline with Physical Priors Annotation; 2.3 RoboScape
- 3 Experiments: 3.1 Experimental Settings; 3.2 Video Quality Evaluation; 3.3 Robotic Policy Learning with Synthetic Data; 3.4 Robotic Policy Evaluation
- 4 Related Work: 4.1 World Model; 4.2 Physics-aware Generative Model; 5 Conclusion and Future Work
- References
- Appendices A Broader Impacts; B Baseline Details; C Scaling Behavior; D.1 Video Generation Results; D.2 Robotic Policy Learning; D.3 Robotic Policy Evaluation
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 reviewed artifact is arXiv:2506.23135v1 [cs.CV], 29 June 2025, marked Preprint. Under review. Its title and all seven authors match the catalog; no other revision was supplied or compared.
- Text extraction alone omits figure images; all twelve figures and three tables were additionally inspected in the supplied PDF.
- Separate supplemental material availability has not been fully verified; no separate supplement or continuous video was reviewed.
- Code was not inspected and experiments were not reproduced. Reference entries were read, but their linked works were not opened.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e01PDF p. 1, title/byline, affiliations and arXiv margin
RoboScape: Physics-informed Embodied World Model; Yu Shang, Xin Zhang, Yinzhou Tang, Lei Jin, Chen Gao, Wei Wu, Yong Li; Tsinghua University and Manifold AI; arXiv v1 dated 29 June 2025.
Go to primary source ↓e02PDF pp. 1–2, Abstract and Section 1
The motivation is expensive robotic data and geometric/motion artifacts in RGB-focused world models; depth and tracked-keypoint supervision are proposed.
Go to primary source ↓e03PDF p. 3, Figure 1 and Section 2.2
The pipeline annotates depth/tracks, slices videos, filters motion/keyframes, and categorizes difficulty/scenes; named preprocessing models are provided.
Go to primary source ↓e04PDF p. 3, Section 2.1, Eq. (1); p. 4, Figure 2 and Section 2.3; p. 5, Eqs. (2)–(4)
Observations are predicted conditional on actions. Frozen-tokenizer symbols, dual branches, temporal/spatial attention, additive conditioning and depth-to-RGB feature fusion define the architecture.
Go to primary source ↓e05PDF p. 5, keypoint dynamics paragraphs and Eqs. (5)–(7); p. 4, Figure 2
Select top-K tracks by accumulated motion; align predicted token features to the predicted initial feature; trajectory locations receive gamma-weighted RGB cross-entropy.
Go to primary source ↓e06PDF p. 6, Eq. (8) and Section 3.1
Provides dataset/task/skill counts, action components, 16 frames at 2 Hz, training clip count, loss weights, epochs, GPU configuration, training duration, and 15-frame inference.
Go to primary source ↓e07PDF p. 7, Table 1, all rows/metric columns, and Section 3.2 metric definitions
RoboScape leads the reported baseline entries for appearance, depth and action sensitivity. CogVideoX action sensitivity is a dash; only qualitative metric definitions are supplied.
Go to primary source ↓e08PDF p. 7, Table 2 and ablation discussion; p. 8, Figure 4 and caption
Ablations trade appearance, depth and action sensitivity. Without keypoints, delta1/delta2 rise to 0.6497/0.8673 from 0.6214/0.8307; qualitative examples highlight local artifacts.
Go to primary source ↓e09PDF p. 8, Table 3, all rows, and Section 3.3 continuation
DP Lift uses synthetic-only training for 10k steps; LIBERO pi0 uses 200 real warm-up trajectories. Table 3 reports separate success-rate curves and real-data comparators.
Go to primary source ↓e10PDF p. 8, Section 3.4; p. 9, Figure 5, panels (a)–(c) and caption
DP checkpoints are evaluated in a post-trained world model and simulator over 100 runs, with manual model-video success labels. Plots report Pearson correlations and R-squared values.
Go to primary source ↓e11PDF p. 9, Section 5, final sentence
Testing the generative world model with real-world robots is future work.
Go to primary source ↓e12PDF p. 6, Section 3.1 Baselines; p. 14, Appendix B
IRASim, iVideoGPT, Genie and CogVideoX are compared; Genie uses an open-source reproduction. Two cited geometric models are excluded for unavailable training code.
Go to primary source ↓e13PDF p. 14, Figure 6 and Appendix C; p. 15, Figure 7 and Appendix C continuation
Model sizes are 34M/131M/544M. The data experiment is described as RoboScape-S at 1000K/3000K/6000K; depth metrics worsen while appearance and action sensitivity improve. Static-frame overfitting is the authors’ explanation.
Go to primary source ↓e14PDF p. 6, Figure 3; p. 8, Figure 4; p. 16, Figures 8–10; p. 17, Figures 11–12
Original frame strips compare predictions and ground truth, including manipulation examples, ablation artifacts, synthetic task data, and policy-evaluation failure/success cases.
Go to primary source ↓8.5 Primary sources
RoboScape: Physics-informed Embodied World Model ↗
PDF · 7,372 extracted words
Source fingerprint
735735cacbf885fc35d75bb1d94eba6b59f4851b85c1ee8a9a976ec387155a45