WorldGym: World Model as An Environment for Policy Evaluation
1. Paper overview
In one sentence: An action-conditioned video simulator can recover useful robot-policy rankings from initial images, but its scores inherit both dynamics and visual-grading errors. e02e05e06e08e09e10e17e18
| At a glance | What to know |
|---|---|
| Research problem | Source description Real-robot evaluation is costly, and manually built simulators require substantial effort. WorldGym asks whether one visual dynamics model learned from diverse offline robot data can estimate the value of multiple policies and tasks using only an initial image and a language goal. e02e21 |
| Core mechanism | |
| A key reported result | Bridge policy-value agreement with real-world trials: WorldGym: RT-1-X 15.5±3.4%; Octo 23.82±4.3%; OpenVLA 67.4±4.9%. Reported r=0.78, p<0.001. Mean task score (%) ± SE across tasks; Pearson correlation. OpenVLA Bridge suite: 17 held-out tasks, 10 matched initial frames per task and policy; original real-world scores versus WorldGym rollouts. Real-world: 18.5±4.0%, 20.0±5.3%, 70.6±6.1%, respectively (Table 5). Aggregate ordering agrees, but this is not per-trial equivalence. Figure 4 rounds RT-1-X to 15.6% and Octo to 23.8%; Table 5 values are preserved here. The approximately 3.3-point average mean gap is in percentage points. e08e09 |
| Reading caution | Reader analysis Realistic object interactions remain difficult. Qualitative control sweeps establish visible action following, not contact-physics accuracy. Figure 17's mean LPIPS stays below 0.2 only over the displayed 20-frame test window; its spread exceeds that threshold, so this is not a trajectory-wise or arbitrary-horizon bound. e07e19e21 |
Core contributions
Figure 1. Actions come from the policy; imagined observations come from WorldGym. Original paper, p. 2 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start at the lower observation o₀ and the yellow goal g. The policy sends actions upward into the blue world-model block, which generates o₁. Follow the returning blue connection to the next policy call, then repeat for o₂ and o₃. These repeated policy boxes depict time steps, not three jointly trained policies. The right-hand VLM grades the generated rollout; although the drawing compresses this connection, Eq. (2) and Appendix B specify the frame sequence and goal as reward inputs. The left-hand variants change the initial image or instruction while retaining the same evaluation loop. e02e03e05e06e21
What it supports. The environment model predicts visual consequences of supplied actions. It does not extract actions from imagined futures to operate the evaluated robot policy. This separation explains how one learned simulator can be reused for multiple policies, and why editing a start image can create a new simulated evaluation setting without constructing a physical scene.
Where the evidence stops. The picture is a system overview, not a DiT layer diagram or evidence of accurate contact physics. The grader and generated observations are learned proxies; the authors explicitly acknowledge imperfect object interactions.
2. Motivation
2.1 The problem and the proposed response
Real-robot evaluation is costly, and manually built simulators require substantial effort. WorldGym asks whether one visual dynamics model learned from diverse offline robot data can estimate the value of multiple policies and tasks using only an initial image and a language goal. e02e21
2.2 What this reading follows
WorldGym turns policy evaluation into an interaction between three distinct components: a robot policy, a video world model and a visual reward grader. Start with a camera image and a task instruction. The policy proposes actions, the world model renders their consequences, and the policy responds to its new synthetic observation. After the rollout, GPT-4o assigns a task score. The strongest evidence comes from reusing initial images from real Bridge trials and comparing aggregate policy scores. Read the six visuals below as separate tests of the loop, its calibration, its grader, and its data and runtime requirements. e02e05e06e08e09e10e17e18
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 | Benchmarks & simulators |
| Architecture | Not applicable |
| Prediction paradigm | Not applicable |
| Quadrant | Not applicable |
3.1 Evidence-based assessment
Supports the recorded classification
The recorded benchmark/simulator classification fits the inference architecture: a learned environment evaluates separate policies using a separate grader. One shared dynamics model does not make this a joint future/action One Model controller. The inverse dynamics module belongs to one evaluated policy, not WorldGym's evaluation mechanism; the recorded architecture, prediction paradigm and quadrant can therefore remain not applicable. e02e03e05e12
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 actor from the learned environment
WorldGym starts from the offline-evaluation problem: estimate how a policy would perform without running each new trial on hardware. The policy still chooses actions from its latest observation and language goal. A separately trained video model then substitutes for the environment's next observation, using action conditioning and causal visual history. Repeating this exchange produces the trajectory that the VLM grader evaluates. This is why the action-conditioned DiT and the language-conditioned video policy in Appendix C must not be conflated. The latter is merely one policy tested inside the evaluator and uses inverse dynamics to produce actions. WorldGym's own world model predicts observations from supplied actions; its reuse across policies depends on that interface separation. e02e03e05e06e12
5.2 Ask which layer of the score has been validated
The main Bridge comparison tests the combined policy, simulator and grader against previously reported real-world scores using matched starting images. Table 3 asks a narrower question: whether the grader recognizes success in real RT-1 videos. These checks are complementary, but neither isolates every possible failure. As reader analysis, a close aggregate mean can arise despite opposing task-level errors, as the Lift White Tape and Put Eggplant into Pot rows demonstrate. Likewise, a reliable real-video grader can behave differently on synthetic artifacts. This distinction matters especially for OOD probes, where edited starts add another changed component. The results support screening and hypothesis generation within the tested setup; they do not establish that each generated success would occur under physical execution. e08e09e10e14e16e21
5.3 Connect the ablations to the evaluator's intended use
The two ablations address different bottlenecks. Bridge V2 improves image-level agreement with validation videos, while matching the prediction horizon to an action chunk reduces generation time. Reader analysis: neither improvement automatically strengthens the central claim about policy evaluation. Better visual metrics may miss object-contact mistakes that change task outcome, and greater parallelism might change the sampled trajectory distribution. A useful reproduction should therefore measure image fidelity, action following and score calibration alongside runtime. The same separation applies to Appendix H: a policy can improve its reward inside WorldGym without a demonstrated gain on real hardware. Treat the preliminary RL curve as evidence that the environment supplies an optimization signal, with transfer and exploitation of model errors still open. e17e18e07e09e20e21
5.4 Training and inference
During training
The 609M-parameter world model has 16 layers, hidden size 1024 and 16 attention heads. It trains on nine compatible Open-X robot datasets with 20-frame context, AdamW at 8e-5, batch size 16 and 300k steps on two A100 80GB GPUs. Table 2 specifies v prediction and a sigmoid noise schedule. e04
Policies are distinct from the evaluator. Appendix C's video policy uses frozen UMT5-xxl conditioning and a separately trained ResNet-50 inverse model with action MSE. Appendix D adapts pretrained Qwen2-VL-2B and a diffusion head using LoRA, training only adapters and the head with noise-prediction MSE. e12e13
During inference
In practice the prediction horizon equals the policy's action chunk size, with ten sampling timesteps and a sliding 20-frame history. OpenVLA uses one-action chunks and Octo four; the latest generated observation supplies feedback only after the chunk. e04e05e18
Monte Carlo evaluation samples policy/world-model rollouts and averages terminal VLM scores. The language goal enters the policy and grader; the dynamics model receives actions and visual context rather than producing goal-directed robot actions itself. e02e03e06
5.5 Implementation flow
- Encode frames and align controls
Stable Diffusion 3's VAE encodes individual 256×256 frames. Seven-dimensional controls combine the paper's six-dimensional end-effector representation and binary gripper state; each component's 10th/90th percentiles are mapped to RT-1's across robots. e04
- Learn action-conditioned dynamics
A latent DiT combines projected action vectors with diffusion-timestep embeddings through AdaLN-Zero. Interleaved spatial and causal temporal attention, trained with Diffusion Forcing, allow autoregressive frame generation. Whole-clip action dropout enables classifier-free guidance. e03
- Close the simulated feedback loop
The external policy consumes the latest frame and goal, emits an action chunk, and receives the final generated frame after the world model predicts one frame per action. Longer rollouts retain a sliding context window; no physical robot executes these simulated actions. e02e05
- Score against the goal
GPT-4o reads generated frames and a task rubric. Scores are binary unless the task permits half credit. Consequently, reported 'success rates' can average graded progress as well as complete task success. e06e09
6. Experiments & results
WorldGym replaces a hand-built evaluation environment with an action-conditioned video model. A separate robot policy repeatedly acts on generated observations; GPT-4o grades the resulting trajectory. Bridge experiments support useful aggregate policy ranking, while task-level discrepancies, imperfect object physics and reward-model transfer limit how confidently simulated scores predict deployment.
6.1 Read the original evidence
Figure 4(a–b). Agreement in overall ranking coexists with scattered task-level estimates. Original paper, p. 5 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the scatter's horizontal axis as real-world success rate and its vertical axis as WorldGym's estimate. Marker shape identifies the policy; the dashed line is a fitted trend, not the perfect-agreement diagonal. The adjacent bars summarize the same comparison by policy, with light bars for real trials and dark outlined bars for generated rollouts. Section 4.1 specifies 17 Bridge tasks and ten matched initial frames per task and policy. For uncertainty, consult Table 5: its standard error is computed across per-task rates, so the bars do not show confidence in every individual physical trial. e08e09
What it supports. The plotted association is r=0.78 with p<0.001, and both environments order the policies RT-1-X, Octo, then OpenVLA. The OpenVLA means are 70.6% in real trials and 67.4% in WorldGym. This supports the use of aggregate scores for this comparison, while the spread of points leaves substantial room for individual task errors.
Where the evidence stops. Figure 4 prints RT-1-X's simulated mean as 15.6%; Table 5 and the main prose give 15.5%. Octo is 23.8% here and 23.82% there. The faithful crop preserves these source differences.
Table 5 and standard-error note. The detailed table reveals errors hidden by similar average scores. Original paper, p. 23 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Within each policy group, compare the real-world column with its WorldGym neighbor for the same row. Every task has ten trials, but several success totals contain halves because some rubrics award partial credit. The bottom row aggregates task scores, and the retained note defines SE as the standard deviation of the task rates divided by the square root of the number of tasks. To see why aggregate agreement is insufficient, locate OpenVLA's Lift White Tape and Put Eggplant into Pot rows: the simulator overestimates the former and underestimates the latter. e06e08e09
What it supports. OpenVLA's Lift White Tape totals are 1 real versus 6 simulated, whereas Put Eggplant into Pot is 10 real versus 6 simulated. Yet its mean remains relatively close: 70.6±6.1% real and 67.4±4.9% simulated. These rows motivate examining task-specific error alongside overall policy ranking, rather than treating the average gap as a universal calibration guarantee.
Where the evidence stops. These real-world scores were taken from the OpenVLA study; WorldGym reuses its initial trial frames. Fractional totals are graded progress, and the across-task SE is not a reported uncertainty bound on simulation bias.
Table 3. The reward check measures grading on real RT-1 videos. Original paper, p. 17 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read this table by column: the left column contains ground-truth successful RT-1 validation videos, and the right contains ground-truth failures. Rows give GPT-4o's success or failure judgment. Thus the upper-right cell measures false positives, not the fraction of predicted successes that are wrong. The caption states that each value is a mean with standard error across four runs. Appendix B explains the instruction-and-rubric prompt that produces the judgment. Keep this validation population separate from Bridge videos generated by the simulator and from scenes initialized through image edits. e06e10e14
What it supports. The reported true-positive rate is 0.81±0.14 and true-negative rate 0.97±0.05. False positives are 0.03±0.05, while false negatives are 0.19±0.14. The low false-positive estimate is encouraging for avoiding overestimated values on these real videos, but missed successes remain another possible source of score distortion.
Where the evidence stops. The number of validation videos per run is not specified. Four-run standard errors do not establish grader reliability under generation artifacts or edited-image distribution shift; that transfer requires a separate labeled test.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| Bridge policy-value agreement with real-world trials OpenVLA Bridge suite: 17 held-out tasks, 10 matched initial frames per task and policy; original real-world scores versus WorldGym rollouts. | WorldGym: RT-1-X 15.5±3.4%; Octo 23.82±4.3%; OpenVLA 67.4±4.9%. Reported r=0.78, p<0.001. Mean task score (%) ± SE across tasks; Pearson correlation | Real-world: 18.5±4.0%, 20.0±5.3%, 70.6±6.1%, respectively (Table 5). Aggregate ordering agrees, but this is not per-trial equivalence. Figure 4 rounds RT-1-X to 15.6% and Octo to 23.8%; Table 5 values are preserved here. The approximately 3.3-point average mean gap is in percentage points. e08e09 |
| Reward-label validation Real RT-1 validation videos with ground-truth success labels; four grader runs. | TP 0.81±0.14; TN 0.97±0.05. Conditional label rates, mean ± SE | FP 0.03±0.05; FN 0.19±0.14. Low false positives support this real-video grading setup. Accuracy on generated or edited scenes remains unvalidated by this table. e10 |
| Policy ranking across versions WorldGym evaluation on the Bridge task suite. | Octo Small/Base 1.5: 21.5/23.8; OpenVLA v0.1/final 7B: 27.6/67.4. Mean reported success score (%) | Larger or later versions receive higher scores; training checkpoints generally improve, with a visible diffusion-policy dip. Version comparisons are consistent with prior rankings cited by the authors; checkpoint improvement is not monotonic or independently validated by new physical trials here. e11 |
| Bridge robustness to unrelated image distractors Edited initial frames across the same Bridge suite; language instructions retained. | RT-1-X 7.6±4.3%; Octo 4.1±1.7%; OpenVLA 39.4±5.1%. Mean simulated success score (%) ± reported uncertainty | Unedited scores: 15.5%, 23.82%, 67.4% in Table 5. All degrade. OpenVLA retains the highest score, but edited-image simulations alone cannot establish physical-world robustness. e09e16 |
| Validation-video fidelity under different training datasets Bridge V1 versus Bridge V2 world-model training, compared with ground-truth validation videos. | Bridge V2: 0.010 / 0.073 / 0.827. MSE ↓ / LPIPS ↓ / SSIM ↑ | Bridge V1: 0.015 / 0.131 / 0.735. All three favor Bridge V2; changing dataset versions does not isolate sample count from composition, nor measure policy-value accuracy. e17 |
| Parallel rollout generation 40-frame generation on an A100 GPU; same world-model checkpoint. | h=4: 33 seconds. Inference time (seconds) | h=1: 93 seconds; reported 2.8× speedup. This supports chunk-aligned parallelism in the measured setting. No paired quality comparison or timing uncertainty accompanies the table. e18 |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Table 8. Bridge V2 training improves all three reported visual-fidelity metrics. Original paper, p. 25 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Compare horizontally within each metric row. The 'Subset (Bridge V1)' column represents the smaller-data training condition and 'Full (Bridge V2)' the larger-data condition. The arrows are essential: lower MSE and LPIPS favor the right column, while higher SSIM also favors it. Appendix F.1 states that generated videos are compared with ground-truth validation videos. This is a diagnostic of visual prediction, not a table of completed robot tasks. Relate it to the evaluator's purpose cautiously: better-looking trajectories are useful, but the ablation does not directly retest whether policy values become more accurate. e17
What it supports. MSE changes from 0.015 to 0.010, LPIPS from 0.131 to 0.073 and SSIM from 0.735 to 0.827. All directions agree in favoring Bridge V2 training. The finding supports investigating the benefit of richer robot data for this video model; it does not quantify a corresponding gain in real-world policy-ranking reliability.
Where the evidence stops. The comparison changes dataset versions, so data composition may change along with size. Appendix F.1 supplies no sample counts, repeated-run uncertainty or matched-compute control sufficient to isolate a pure scaling effect.
Table 9. Chunk-aligned frame denoising reduces the measured rollout-generation time. Original paper, p. 25 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Here h means frames predicted in parallel, not the total length of the task. The caption fixes the workload at a 40-frame video rollout on an A100 GPU. Compare h=1 with h=4 while retaining the same world-model checkpoint. Appendix F.2 connects these settings to actual policy interfaces: OpenVLA emits a single action and Octo emits four. The broader inference rule is to set h to the policy's action chunk length. That lets the model process already available actions together; it does not create future actions that a feedback policy has yet to decide. e05e18
What it supports. Generation takes 93 seconds at h=1 and 33 seconds at h=4, the paper's reported 2.8× speedup. This directly supports parallel denoising as an efficiency mechanism for policies that provide action chunks. It is a measured result for one workload and GPU, rather than a throughput guarantee for every policy or rollout configuration.
Where the evidence stops. The table reports time without a paired quality comparison or timing uncertainty. Increasing a policy's own action chunk can also change when it receives feedback, so a controlled timing test should keep its supplied actions fixed.
7. Analysis & limitations
7.1 What the evidence leaves open
Realistic object interactions remain difficult. Qualitative control sweeps establish visible action following, not contact-physics accuracy. Figure 17's mean LPIPS stays below 0.2 only over the displayed 20-frame test window; its spread exceeds that threshold, so this is not a trajectory-wise or arbitrary-horizon bound. e07e19e21
Good aggregate agreement can hide task errors: OpenVLA scores 1 versus 6 on Lift White Tape and 10 versus 6 on Put Eggplant into Pot in real versus simulated trials, each out of ten. Partial credit further separates reported score from strict binary completion. e06e09
OOD behavior mixes policy, image-editing, dynamics and grader errors. Table 6 calls one edit 'Replace Carrot with Radish', while Figure 9 describes recoloring the carrot. Table 1 omits explicit OOD-language trial denominators; no rates are inferred from it. e14e15
Appendix H's RL improvement is measured inside WorldGym, with an initial dip, and supplies no real-world transfer test. Google Robot results likewise lack paired real-world columns and allow half credit for reaching attempts. e20e22
7.2 Questions for discussion
- Would policy ordering persist when humans grade the same generated videos and when difficult tasks are evaluated separately?
- Does chunk-aligned denoising preserve action fidelity and policy-value calibration while reducing runtime?
8. Reproducibility audit
8.1 Requirements and known gaps
A faithful implementation needs the exact nine-dataset mixture/splits, action transforms, SD3 VAE, DiT checkpoint and policy versions. The paper supplies percentile alignment and core hyperparameters but omits the complete mixture, action-dropout probability, guidance scale, software versions and whether the VAE is frozen. e03e04
Recreate matched initial frames, rollout horizons and per-task rubrics; pin the GPT-4o version and frame-selection procedure. Those grader details and validation sample counts are incomplete. Algorithm 1 also leaves n_chunk undefined and drops the goal in its terminal reward call, unlike Eq. (2) and the prompt. e02e05e06e08e10
Checkpoint reproduction needs clarification: Section 4.2 lists video checkpoints ending at 18k steps, whereas Appendix C reports 180k generator training. The main text's 'from scratch' description also differs from Appendix D's pretrained adaptation. Neither discrepancy is silently resolved here. e11e12e13
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 grader error from simulated-outcome error
Proposed check, not performed: reuse the matched Bridge initial frames and fixed policy versions, then collect repeated WorldGym rollouts with recorded seeds. Have blinded humans apply the same binary/partial-credit rubrics to both generated and available real videos, and have a pinned GPT-4o version score those same clips. Compare VLM–human disagreement by video type, then compare human-graded simulated versus real task scores. Report task-level errors and policy rankings, with task-resampled uncertainty. Repeat on a prespecified distractor subset. A larger VLM–human gap on generated clips would implicate grader transfer; persistent real/simulated disagreement after human grading would instead implicate the simulated outcomes. e06e08e09e10e16
Check 2: Test parallelism without changing the policy's decisions
Proposed check, not performed: fix initial images, world-model checkpoint, sampling settings and complete 40-action sequences, then generate the same workload with h=1 and h=4 on the same A100. Repeat matched runs and measure latency, LPIPS against available recorded trajectories, visible action-following errors and blinded task scores. Use already fixed actions so the h=1 condition cannot replan more often. Afterwards, repeat interactive Octo rollouts while keeping its four-action policy calls unchanged and varying only how each chunk is rendered. The efficiency claim is stronger if runtime falls without a consistent fidelity or scoring loss; score shifts at fixed actions would expose a quality tradeoff absent from Table 9. e04e05e07e18e19
8.3 Reading coverage
Visual audit: Visually inspected the title/author page, all main-method and experiment pages, and every appendix page; this includes Figures 1–18, Tables 1–9, Algorithm 1, hyperparameters, reward rubrics, policy-training details and the RL/drift diagnostics. All six final original-PDF crops were separately viewed and checked for complete labels, legible numbers and preserved source pixels. The Table 5 crop retains its standard-error definition. Figure 1's feedback direction was checked against Sections 2–3 and Algorithm 1; the diagram compresses the sequence/goal inputs to the grader, which are explicit in Eq. (2) and Appendix B. Printed rounding differences and pseudocode/training ambiguities are disclosed in the base report. Reference-only pages 10–15 were read as text; separate supplemental videos, code and other sources were not inspected.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 9, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26. 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 and Sections 1–2: motivation and policy-value formulation
- Sections 3.1–3.2: world-model training, rollout, reward and validation
- Sections 4.1–4.3: real-world comparison, ranking and OOD probes
- Sections 5–6, acknowledgments and references
- Appendix A: architecture, hyperparameters and Algorithm 1
- Appendix B.1–B.2: reward prompt and validation
- Appendices C–D: evaluated video and diffusion policies
- Appendix E.1–E.4: additional robot and OOD results
- Appendix F.1–F.2: dataset and parallelism ablations
- Appendices G–H: visual drift and preliminary RL
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.
- All seven supplied text chunks were read individually, covering all 26 PDF pages. Original PDF figures and tables were inspected separately because text extraction does not reconstruct images.
- Separate supplemental material availability has not been fully verified. Linked videos, code, external cited papers and other online resources were not inspected; experiments were not reproduced.
- Identity scope: the supplied ICLR 2026 proceedings PDF matches the catalog title and all six authors. The acquisition title merges some interword spaces; its exact observedTitle is preserved in provenance and metadata. No numbered revision or revision date is established. No comparison with an earlier OpenReview artifact was possible from the supplied material.
- Visual inspection covers PDF pp. 1–9 and 16–26, including every figure and table. Reference-only pp. 10–15 were read as text.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e01PDF p. 1, title, author block and conference header
The supplied artifact is titled WorldGym: World Model as An Environment for Policy Evaluation and marked as an ICLR 2026 conference paper. Authors, in order, are Julian Quevedo, Ansh Kumar Sharma, Yixiang Sun, Varad Suryavanshi, Percy Liang and Sherry Yang. Affiliations are Stanford University, NYU and Google DeepMind. No numbered revision or revision date is printed here.
Go to primary source ↓e02PDF pp. 2–3, Figure 1 and Section 2, Eqs. (1)–(2)
WorldGym estimates policy value by sampling visual trajectories from an action-conditioned world model and scoring them with a learned reward function. The goal and initial state are sampled from G; observations replace inaccessible physical states in the evaluation loop. Figure 1 depicts separate policy, world-model and reward components.
Go to primary source ↓e03PDF p. 3, Section 3.1.1
The latent Diffusion Transformer uses Diffusion Forcing, linear action projection added to diffusion-timestep embeddings, AdaLN-Zero conditioning, clip-level action dropout and classifier-free guidance. Causal temporal attention alternates with spatial attention.
Go to primary source ↓e04PDF p. 16, Appendix A, implementation paragraph and Table 2
Stable Diffusion 3's VAE encodes 256×256 frames. The world model uses 609M parameters, 16 layers, hidden size 1024, 16 heads, 20-frame context and seven action dimensions. Nine compatible Open-X datasets are used; action-component 10th/90th percentiles are aligned to RT-1. Training uses AdamW, learning rate 8e-5, batch size 16, 300k steps and two A100 80GB GPUs; Table 2 lists a sigmoid noise schedule, v prediction and 10 sampling timesteps. The complete dataset mixture, guidance scale and dropout rate are not specified here.
Go to primary source ↓e05PDF pp. 3–4, Section 3.1.2; p. 16, Algorithm 1 and following paragraph
The policy receives the latest observation and emits an action chunk; the world model generates a frame per action and returns the latest frame. In practice h equals the action chunk size. History uses a sliding 20-frame window. Algorithm 1 increments n by an undefined n_chunk and writes the terminal reward without its goal argument, unlike Eq. (2).
Go to primary source ↓e06PDF p. 4, Section 3.1.3; p. 17, Appendix B.1, reward prompt and rubrics
GPT-4o scores the generated frame sequence against the language instruction. Binary tasks use 0 or 1; tasks with a specified partial-credit criterion allow 0.5. The prompt acknowledges generation artifacts. An exact GPT-4o snapshot, frame-sampling specification and all task-specific partial-credit descriptions are not supplied.
Go to primary source ↓e07PDF pp. 4–5, Sections 3.2.1–3.2.2 and Figures 2–3; p. 21, Figures 15–16
Validation videos are generated from initial frames and the original recorded actions, with qualitative comparisons across robot morphologies. Separate control sweeps change one action dimension while holding others at zero. The stills illustrate action following; the paper directs readers to supplemental videos for motion.
Go to primary source ↓e08PDF pp. 5–6, Section 4.1 and Figure 4
The evaluation reuses initial frames from OpenVLA's real-world Bridge trials: 17 tasks absent from Bridge V2, 10 trials per task and policy. RT-1-X, Octo and OpenVLA are rolled out in WorldGym. Figure 4 reports Pearson r=0.78, p<0.001; its bars show real/simulated means of 18.5/15.6%, 20.0/23.8% and 70.6/67.4%. Page 6 instead gives 15.5% and 23.82% for the first two simulated means.
Go to primary source ↓e09PDF p. 23, Appendix E.3, Table 5, all rows and standard-error definition
Table 5 reports real/WorldGym means (±SE): RT-1-X 18.5±4.0% / 15.5±3.4%; Octo 20.0±5.3% / 23.82±4.3%; OpenVLA 70.6±6.1% / 67.4±4.9%. SE is sd of per-task rates divided by the square root of the number of tasks. OpenVLA's Lift White Tape scores are 1 real and 6 simulated; Put Eggplant into Pot scores are 10 real and 6 simulated, each over 10 trials. Fractional success totals occur in the table.
Go to primary source ↓e10PDF pp. 17–18, Appendix B.2; p. 17, Table 3
On real RT-1 validation videos with ground-truth success labels, GPT-4o has TP 0.81±0.14, FN 0.19±0.14, FP 0.03±0.05 and TN 0.97±0.05. The caption identifies mean and standard error across four runs. It does not specify the number of videos per run or establish accuracy on generated OOD videos.
Go to primary source ↓e11PDF pp. 6–7, Section 4.2 and Figures 6–7
Figure 6 gives WorldGym scores of 21.5%, 23.8%, 27.6% and 67.4% for Octo Small 1.5, Octo Base 1.5, OpenVLA v0.1 7B and OpenVLA 7B. Figure 7 shows an overall upward checkpoint trend, with a diffusion-policy decline from 20k to 40k. Page 7 describes video-policy evaluations at 2k, 8k, 12k and 18k steps, and calls both trained policies 'from scratch'.
Go to primary source ↓e12PDF p. 19, Appendix C and Figure 14
The evaluated video policy is a separate UniPi-style language-conditioned generator and inverse dynamics model. Its frozen UMT5-xxl encoder provides pooled language features; a ResNet-50/MLP inverse model uses frame features and temporal differences, trained with action MSE. Appendix C states 180k generator steps and 200k independent inverse-model steps, leaving the relation to the main text's 18k checkpoint endpoint unclear.
Go to primary source ↓e13PDF p. 20, Appendix D
The diffusion policy uses pretrained Qwen2-VL-2B and a pretrained control head. Only LoRA adapters and the policy head are trained during 60k-step Bridge V2 adaptation; actions are rescaled to (-1,1). The head uses noise-prediction MSE and DDIM inference. This pretrained adaptation qualifies the main text's 'from scratch' description.
Go to primary source ↓e14PDF pp. 7–9, Section 4.3, Figures 8–12 and Table 1
Image edits introduce objects, distractors and classification targets; changed instructions alter objects or destinations. Table 1 lists scores for four OOD language tasks but omits an explicit trial-count column or rate unit. Figure 9 describes turning the carrot red; this differs in wording from Table 6's radish replacement. These are simulated policy probes.
Go to primary source ↓e15PDF p. 24, Appendix E.4, Table 6
OpenVLA's orange-selection scores are 4/10 with carrot closer, 9/10 with orange closer and 10/10 for 'Replace Carrot with Radish'. Its classification totals are 20/20 for red/blue, 12/20 for circle/square and 11/20 for the two celebrity faces. The table does not establish equivalent behavior on real edited scenes.
Go to primary source ↓e16PDF pp. 8–9, Section 4.3 and Figure 13; p. 24, Table 7
Unrelated distractors are added to initial images across the Bridge suite while instructions are held fixed. Table 7 reports mean simulated success scores of 7.6±4.3%, 4.1±1.7% and 39.4±5.1% for RT-1-X, Octo and OpenVLA. Figure 13 compares them with the unedited-image scores.
Go to primary source ↓e17PDF p. 25, Appendix F.1, Table 8 and following paragraph
Bridge V1 versus Bridge V2 training yields MSE 0.015 versus 0.010, LPIPS 0.131 versus 0.073 and SSIM 0.735 versus 0.827 against ground-truth validation videos. Lower is better for MSE/LPIPS and higher for SSIM. Validation sample counts, repetitions and uncertainty are not reported in this ablation.
Go to primary source ↓e18PDF p. 25, Appendix F.2, Table 9 and following paragraph
Generating 40-frame rollouts on an A100 takes 93 seconds at h=1 and 33 seconds at h=4, a reported 2.8× speedup. The same checkpoint can match policy chunk sizes: OpenVLA predicts one action, Octo four. The table provides timing, without paired quality or repeated-timing uncertainty.
Go to primary source ↓e19PDF p. 26, Appendix G and Figure 17
Mean per-frame LPIPS over ten test rollouts stays below 0.2 under fixed recorded actions. The plotted frame indices span 0–19 and the shaded band is ±1 standard deviation, extending above 0.2 late in the rollout. Error grows over time; a bound on every trajectory or on hundreds of steps is not demonstrated.
Go to primary source ↓e20PDF p. 26, Appendix H and Figure 18
A preliminary RL4VLA study finetunes OpenVLA inside WorldGym from Bridge V2 training starts and evaluates with Bridge V2 test starts and instructions inside WorldGym. The displayed reward trend improves overall but dips initially. No real-world transfer result is reported for this RL study.
Go to primary source ↓e21PDF p. 1, abstract; p. 9, Section 6
The authors acknowledge that highly realistic object interactions remain challenging and that not every generated interaction is fully realistic. They position WorldGym as a preliminary evaluation aid before deployment.
Go to primary source ↓e22PDF p. 22, Appendix E.2, Table 4
Four Google Robot tasks from RT-1 use ten trials each and give half credit for attempting to reach the target location. On Pick Blue Chip Bag, the reported totals are 5 for RT-1-X, 5.5 for Octo and 9 for OpenVLA. Table 4 compares simulated policies without paired real-world evaluation columns.
Go to primary source ↓8.5 Primary sources
WORLDGYM: WORLDMODEL ASANENVIRONMENT FORPOLICYEVALUATION ↗
PDF · 11,976 extracted words
Source fingerprint
747d71a0b1e79b5dedb77c79dbe51d3a8ecaab36d6eadf5115ef3085d3d0a9a5