PAPER REPORTENAll readings ↗

Genie Envisioner: A Unified World Foundation Platform for Robotic Manipulation

English reading report: Method, equations, original figures, experiments and reproducibility.

Authors: Yue Liao; Pengfei Zhou; Siyuan Huang; Donglin Yang; Shengcong Chen; Yuxin Jiang; Yue Hu; Jingbin Cai; Si Liu; Jianlan Luo; Liliang Chen; Shuicheng Yan; Maoqing Yao; Guanghui Ren

Affiliations: AgiBot Genie Team; LV-NUS Lab; BUAA

Source: ICLR 2026 · 2508.05635 ↗ · Project page ↗ · Catalog record

Reading: 340 / 558 · 6 original figures & tables · ~22 min ·

1. Paper overview

In one sentence: Genie Envisioner turns robotic video priors into fast action decoding and action-conditioned simulation, but its transfer requires supervision and its simulation evidence remains based on video metrics. ev-platformev-actionev-action-trainingev-sim-conditioningev-ablationev-sim-results

At a glanceWhat to know
Research problem
Source description

Robot policy learning and evaluation typically require separate infrastructure. The authors propose sharing a generative representation of robot–scene dynamics across prediction, control and simulation. The practical question is whether video pretraining supplies useful control features while retaining enough action sensitivity for evaluation. ev-platform

Core mechanism
Source description

GE-Base adapts a video diffusion transformer to synchronized head and wrist views, language instructions and sparse historical observations. Selected blocks exchange information across views. ev-base

A key reported resultRed cylinder into a paper cup: GE-Base initialization plus task-specific video adaptation: E2E 0.89; SR 0.76.

End-to-end success (E2E); step-wise success (SR). AgiBot G1; fixed positions; 305 demonstrations; 40,000 training steps per model; robot state included. Evaluation trial count is unstated.

GE-Base initialization without adaptation: E2E 0.81; SR 0.64. Gains are 8 and 12 percentage points. Without state, adaptation lowers E2E from 0.49 to 0.37. E2E permits retries; SR counts completed sub-steps. No uncertainty is reported. ev-ablationev-robot-eval

Reading caution
Source description

Authors identify a single-platform robotic pretraining corpus, parallel-jaw upper-body manipulation, and proxy evaluation with partial human validation. Dexterous hands, locomotion and broad heterogeneous-source robustness remain untested. ev-limitations

Core contributions

  • Source description

    GE-Base adapts a video diffusion transformer to synchronized head and wrist views, language instructions and sparse historical observations. Selected blocks exchange information across views. ev-base

  • Source description

    GE-Act adds a separate 160M-parameter action transformer; GE-Sim instead adds pose-image and motion conditioning. These are distinct uses of a shared foundation, with different inference outputs. ev-actionev-sim-interfaceev-sim-conditioning

  • Source description

    EWMBench combines scene, motion and semantic metrics on manipulation tasks held out from robotic pretraining, plus a limited comparison with human video rankings. ev-bench-dataev-bench-metricsev-human

Figure 6. GE-Act attaches a smaller action pathway to the visual backbone at corresponding transformer depths. Original paper, p. 8 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. Start at the image grid and instruction on the left. Their encoders feed the blue GE-Base column, whose successive visual DiT blocks produce intermediate features. Follow the horizontal arrows into the green action blocks: these indicate the visual-to-action conditioning described as cross-attention in Section 3.1. The action pathway starts from noisy actions, follows the visual backbone’s depth with a smaller hidden width, and ends in action tokens. The purple video-token output is a latent representation; the method does not require decoding a future video to control the robot. The right-hand robot examples indicate deployment targets, with embodiment-specific adaptation described separately. ev-actionev-action-trainingev-fastev-transfer

What it supports. The source describes a 160M-parameter action branch that can reuse a much larger video model’s features. With the asynchronous procedure on page 10, one visual denoising step supplies cached features to five action denoising steps; the reported 54-step forward pass takes 200 ms on an RTX 4090.

Where the evidence stops. The separate branches support the Dual-system assessment despite joint task fine-tuning. The pictured new embodiments require demonstrations and a newly trained action head; the diagram alone establishes neither zero-shot transfer nor robot success rates.

2. Motivation

2.1 The problem and the proposed response

Source description

Robot policy learning and evaluation typically require separate infrastructure. The authors propose sharing a generative representation of robot–scene dynamics across prediction, control and simulation. The practical question is whether video pretraining supplies useful control features while retaining enough action sensitivity for evaluation. ev-platform

2.2 What this reading follows

A robot needs both an action to execute and a useful account of what that action will do. Genie Envisioner develops these capabilities around GE-Base, a video model adapted to robotic observations and instructions. GE-Act reads intermediate visual features to produce control trajectories, while GE-Sim takes trajectories as inputs and generates their visual consequences. Reading the architecture alongside its training schedule explains why deployment can skip explicit video decoding. Reading the experiments requires a second distinction: executed robot success, generated-video quality and simulator fidelity are different measurements. The figures below expose both the advantages of embodied pretraining and the qualifications hidden by broad platform-level claims. ev-platformev-actionev-action-trainingev-sim-conditioningev-ablationev-sim-results

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 dimensionRecorded classification
Major categoryWAMs
ArchitectureDual-system
Prediction paradigmOther mechanisms
QuadrantOutside quadrants

3.1 Evidence-based assessment

Supports the recorded classification

Reader analysis

Dual-system / Other mechanisms / Outside quadrants is supported by GE-Act’s separate visual and action transformers, coupled through blockwise cross-attention and cached features. It decodes actions from predictive latents without rendering a future video for inverse dynamics. Joint fine-tuning does not establish one shared transformer jointly emitting future pixels and actions. GE-Sim is a distinct action-to-video simulator. Embodiment adaptation and asymmetric inference support the recorded subcategories. ev-actionev-action-trainingev-fastev-transferev-sim-loop

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

InputsOutputs
  • Language instruction, initial synchronized head/left-wrist/right-wrist images, and sparse historical frames
  • Robot state in the Table 1 state-conditioned policy variants
  • GE-Sim: action trajectories, calibrated cameras and a reference image
  • GE-Base: autoregressive multi-view video chunks
  • GE-Act: executable action trajectories
  • GE-Sim: action-conditioned video chunks, optionally fed back to a policy

4.2 Equations and their role

x1:N(t)=W(x^0:t1,x0,q)x_{1:N}^{(t)}=\mathcal{W}(\hat{x}_{0:t-1},x_0,q)
World model W generates an N-frame chunk at step t from sparse history, initial observation x₀ and instruction q. Here N counts frames; the attention discussion later reuses N for camera views. ev-base
vi=E(Ii)+E(Pi),Δai=aiai1=[Δpi,Δri]v_i=\mathcal{E}(I_i)+\mathcal{E}(P_i),\qquad \Delta a_i=a_i-a_{i-1}=[\Delta p_i,\Delta r_i]
GE-Sim encoder E maps history frame Iᵢ and projected pose image Pᵢ to features whose sum is vᵢ. For motion conditioning, aᵢ contains position pᵢ and orientation rᵢ; consecutive differences define motion input. These are Eqs. (1)–(2). ev-sim-conditioning
SAscore=1dsymH(G,P)+ϵ\mathrm{SA}_{\mathrm{score}}=\frac{1}{d_{\mathrm{symH}}(G,P)+\epsilon}
Spatial alignment inverts symmetric Hausdorff distance between reference trajectory G and detected generated trajectory P. Higher is better. The source sets ε to 10⁻⁸ and retains the best spatial match among three generated samples for further trajectory evaluation. ev-bench-metrics

5. Method in detail

5.1 Learn future-oriented features, then read out control

Source description

The foundation model first learns to predict robotic video chunks from instructions, initial observations and sparse history. Its three camera views are encoded into a shared latent format, with selected cross-view attention blocks exchanging information. GE-Act then attaches a smaller action pathway to intermediate features at corresponding depths. Action pretraining holds the visual backbone fixed and supervises only action trajectories, so useful control information must be available in the pretrained visual representation. At deployment, the model does not need to turn the predicted latent features into displayable video. A single visual denoising step produces cached features for five action denoising steps. This explains the mechanism behind the reported low trajectory-generation latency; it does not by itself establish how accurately those features model counterfactual outcomes. ev-baseev-actionev-action-trainingev-fast

Figure 7. The action model is trained after the foundation model, then adapted through separate video and action phases. Original paper, p. 9 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. Read the three rows from top to bottom, keeping the flame symbols and rightmost losses in view. Stage 1 trains the action DiT using action supervision while GE-Base-LF is fixed. Stage 2 updates the video DiT with video supervision; Section 3.2 specifies a mixture of the full corpus and task data weighted tenfold. Stage 3 updates both the visual backbone and action module using task-specific action supervision. The left-hand streams explain which data enter adaptation, while the 5 Hz and 30 Hz labels distinguish visual sampling from action resolution. These are post-foundation stages: the earlier MR and LF video pretraining is described on page 6. ev-base-trainingev-action-trainingev-training-discrepancy

What it supports. The recipe separates learning a reusable visual representation from learning an embodiment’s control mapping. Action pretraining is reported as three days on 16 A100 GPUs, followed by 12-hour video adaptation on eight A100s. Small task-demonstration requirements therefore coexist with substantial prior training; they are not the total cost of learning the platform.

Where the evidence stops. The caption says the visual backbone is optimized first and the video encoder adapted later, conflicting with the flame markers and Section 3.2. Stage 3 also says 24 hours here versus 36 hours in the prose. The explanation follows Section 3.2 while leaving those implementation discrepancies unresolved.

5.2 Read adaptation as an interaction with state and embodiment

Reader analysis

Table 1 provides a more discriminating question than whether pretraining helps on average. Holding the state input fixed, GE-Base initialization produces much better success than the row without embodied initialization. Adding task-specific video adaptation improves both E2E and SR when state is available. Without state, however, E2E drops while SR rises. Reader interpretation: the adaptation benefit depends on the information available to the policy and on the success definition, so the best row cannot justify a universal claim. The transfer protocol introduces another distinction. On Agilex and Dual Franka, the visual model is adapted and a new action head is trained from scratch because action spaces differ. The approximately one-hour demonstrations measure supervised adaptation effort on each task, not zero-shot competence or the total pretraining budget. ev-ablationev-robot-evalev-transferev-base-training

Figure 14. GE-Sim supplies action information through both projected geometry and temporal motion features. Original paper, p. 16 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. Trace the action sequence in the left panel into two paths. The lower path projects end-effector positions, orientation axes and gripper openness into camera-aligned pose images. These and corresponding history frames share the video encoder, and their latents are added according to Eq. (1). The upper path computes consecutive pose differences, encodes them, and brings them together with a CLIP reference-image style token for cross-attention. Noise and the visual condition enter the generative model, whose decoder produces video frames. In the right panel, follow Act downward from the policy to GE-Sim and generated observations back toward the policy; the lower arrows describe initializing scenes and augmenting trajectories. ev-sim-interfaceev-sim-conditioningev-sim-symbolev-sim-loop

What it supports. The simulator receives a commanded trajectory rather than choosing that trajectory itself. This lets a policy interact with generated visual observations and lets the same trajectory be reused under different starting imagery. That architectural capability motivates closed-loop evaluation, but its usefulness depends on whether generated consequences remain faithful when actions deviate from demonstrations.

Where the evidence stops. The upper merge is drawn as a circled plus, although both caption and text specify concatenation of motion and reference features. The lower plus agrees with Eq. (1)’s addition. The crop preserves both symbols; the written concatenation rule guides the upper-path explanation.

5.3 Separate action faithfulness from an attractive simulated future

Reader analysis

GE-Base predicts plausible instruction-following futures, whereas GE-Sim must respect a supplied action sequence. Pose2Image provides camera-aligned geometry, and motion deltas provide changes across time; a reference-image token helps anchor appearance. Feeding generated observations back into a policy creates a closed loop in which simulator errors could change later actions. Reader interpretation: that feedback makes action faithfulness more demanding than producing good videos under known trajectories. Table 2 evaluates ground-truth-action-conditioned generation and shows different winners for dynamic consistency and spatial alignment. EWMBench also selects the best spatial trajectory from three samples and uses learned semantic judges. These measurements are useful diagnostics, but they do not establish that a simulator will preserve real-world policy rankings under failed grasps, contacts or recovery behaviors. ev-baseev-sim-conditioningev-sim-loopev-sim-resultsev-bench-metricsev-limitations

5.4 Training and inference

During training

Source description

GE-Base uses approximately one million AgiBot-World-Beta episodes, totaling 2,967 hours. MR training samples 57 frames at 3–30 Hz plus four memory frames, producing eight latent frames; LF adaptation uses nine frames at 5 Hz plus four memory frames, producing two latent frames. The VAE remains frozen. Compute is 32 A100 GPUs for seven days, then three days. ev-base-training

Source description

GE-Act pretraining freezes GE-Base-LF and trains only the action module on trajectory supervision: four conditioning frames at 5 Hz and 54 action steps at 30 Hz. Video generation is disabled. Adaptation first updates video generation using the full corpus plus task data weighted 10×, then specializes backbone and action module on task data alone. Figure 7 and prose disagree about updates and duration. ev-action-trainingev-training-discrepancy

Source description

New embodiments receive video adaptation followed by an action head trained from scratch. GE-Sim initializes from GE-Base-MR, uses ground-truth actions and additional failed/suboptimal trajectories, freezes VAE and CLIP encoders, and optimizes remaining parameters with a video flow-matching loss. ev-transferev-sim-training

During inference

Source description

GE-Base uniformly samples historical frames at inference. GE-Act caches visual features after one denoising step and reuses them for five action denoising steps. The stated 5 Hz visual processing and 30 Hz action resolution are distinct from complete trajectory-prediction throughput. ev-memoryev-fast

Source description

In GE-Sim’s closed loop, a policy proposes actions and receives a generated video chunk as its next observation. Replaying actions under changed initial imagery is proposed as a data engine. This is a learned visual environment rather than measured physical execution. ev-sim-loop

5.5 Implementation flow

  1. Construct a predictive visual representation

    A shared video encoder maps initial and historical observations to tokens. View embeddings, spatial positions and noise accompany these tokens; frozen T5-XXL embeddings supply language through cross-attention. Selected DiT blocks exchange features across views. The paper uses LTX-Video 2B for fast control and also considers COSMOS2 2B for simulation. ev-base

  2. Decode actions from intermediate features

    A narrower action DiT mirrors the visual transformer’s depth. Each action block cross-attends to corresponding visual features while refining noisy action tokens. The trajectory can be executed without decoding a future video; new physical observations provide feedback. ev-actionev-action-training

  3. Condition a simulator on commanded motion

    GE-Sim represents each arm by end-effector position, orientation and gripper openness. Camera calibration projects poses into images. Encoded pose images are added to encoded history frames; separately encoded pose differences and reference-image features condition the DiT through cross-attention. ev-sim-interfaceev-sim-conditioning

6. Experiments & results

Genie Envisioner transfers robotic video-generation representations into two downstream systems: GE-Act decodes executable action chunks from visual latents, while GE-Sim predicts videos conditioned on supplied actions. Its controlled ablation supports embodied pretraining, with state-dependent benefits from task adaptation. Cross-embodiment results require demonstrations and a new action head; simulator video scores do not establish physical policy success.

6.1 Read the original evidence

Figure 17. GE-Base’s largest visible advantage is in the motion aggregate under instruction-and-image-conditioned generation. Original paper, p. 20 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. Start with the table rather than comparing radar areas. The table groups the benchmark into scene, motion and semantics, then lists a reported total; these are composite scores rather than success probabilities. The radar unpacks spatial alignment, temporal alignment, dynamic consistency, scene consistency, logic, CLIP, BLEU and diversity. Its color keys align with the adjacent model rows. Section 6 evaluates language- and head-image-conditioned videos on ten pretraining-disjoint tasks with 100 instances each. For trajectory evaluation, the best spatial match among three generated samples is retained. This selection protocol matters when interpreting motion quality as the reliability of a typical generated rollout. ev-bench-dataev-bench-metricsev-bench-results

What it supports. GE-Base scores 4.7010 overall versus Kling’s 3.8698. Its motion aggregate is 1.6676 versus 0.9440, whereas semantics are closer at 2.0907 versus 2.0370. The observed pattern supports improved robotic motion modeling after domain adaptation, within the benchmark’s reported scoring system and selected generation protocol.

Where the evidence stops. Six models appear in the graphic, although Section 6.3 says seven; no missing result is inferred. Domain pretraining differs across models. The scores concern generated videos and proxy evaluators, so they cannot be read as physical task-success rates or matched-training comparisons.

Table 2. Changing the simulator backbone improves some fidelity metrics while reducing others. Original paper, p. 21 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. Read vertically within each column, keeping both metric blocks together. The row labeled COSMOS refers to the COSMOS2-based GE-Sim variant in the caption. BLEU and CLIP assess semantic alignment; DYN compares motion dynamics, and Div. records generated diversity. The second block contains spatial alignment, logical consistency, temporal alignment and scene consistency. The scores have different scales, so a larger number in one column is not evidence that it matters more than another metric. Section 6.4 supplies ground-truth actions to both simulators. Under fixed actions, the authors interpret low diversity as consistent action-to-video correspondence, unlike open-ended generation where diversity can be desirable. ev-bench-metricsev-sim-resultsev-sim-loopev-limitations

What it supports. COSMOS2 improves DYN from 0.78 to 0.85 and reported PSNR from 19.9 to 20.7. LTX instead leads on SA, 0.94 versus 0.87, and TA, 0.98 versus 0.97; logic is 0.97 for both. The table therefore supports a mixed backbone tradeoff rather than uniform COSMOS2 superiority.

Where the evidence stops. These are videos conditioned on ground-truth control, not success rates from policies interacting with the simulator or real robots. Preserve the reported scales: the table does not fully specify CLIP scaling or the PSNR evaluation protocol. Closed-loop error accumulation remains a separate question.

6.2 Results and evaluation conditions

Task & protocolReported resultComparison & interpretation
Red cylinder into a paper cup

AgiBot G1; fixed positions; 305 demonstrations; 40,000 training steps per model; robot state included. Evaluation trial count is unstated.

GE-Base initialization plus task-specific video adaptation: E2E 0.89; SR 0.76.

End-to-end success (E2E); step-wise success (SR)

GE-Base initialization without adaptation: E2E 0.81; SR 0.64.

Gains are 8 and 12 percentage points. Without state, adaptation lowers E2E from 0.49 to 0.37. E2E permits retries; SR counts completed sub-steps. No uncertainty is reported. ev-ablationev-robot-eval

GE-Act trajectory inference latency

Onboard NVIDIA RTX 4090; cached visual features and asynchronous denoising.

Reported 200 ms.

Forward-pass latency for a 54-step action trajectory

One visual denoising step supplies features to five action denoising steps; no matched latency baseline is tabulated.

A trajectory-generation measurement, not a sensing-to-actuation latency distribution. ev-fast

Cross-embodiment cloth and box folding

Agilex: cloth and box folding; Dual Franka: cloth folding. Each task uses 250 demonstrations, approximately one hour, for adaptation.

GE-Act has higher plotted SR and E2E than the compared VLA baselines; exact bar values are unlabeled.

SR and E2E in Figure 11

GR00T N1, UniVLA and π₀ receive the respective same adaptation datasets.

Supports supervised transfer with a new action head, not zero-shot transfer or matched pretraining budgets. ev-transfer

Four-task RoboTwin adaptation

One GE-Act model jointly adapted on 200 demonstrations, 50 per task; baselines adapted separately per task.

Authors report better performance on three of four tasks, with lower performance on lift pot.

Task success in Figure 11(c)

π₀ and GO-1 under task-specific adaptation.

The multitask setting differs; the suggested explanation of task interference is not isolated experimentally. ev-robotwin

Instruction-conditioned video generation on EWMBench

Ten pretraining-disjoint AgiBot test tasks, 100 instances each; language plus head-view image conditioning.

GE-Base: 4.7010.

Aggregate score in Figure 17(b)

Kling: 3.8698; unadapted LTX: 2.9676.

A video-quality score on the reported scale, not physical success. GE-Base receives robotic domain pretraining. ev-bench-dataev-bench-results

Ground-truth-action-conditioned GE-Sim evaluation

EWMBench simulator comparison supplied ground-truth control trajectories.

COSMOS2: DYN 0.85; SA 0.87; PSNR 20.7.

Dynamic consistency (DYN); spatial alignment (SA); PSNR, as reported

LTX: DYN 0.78; SA 0.94; PSNR 19.9.

COSMOS2 improves DYN and PSNR while LTX has higher SA. The table does not measure autonomous closed-loop policy success. ev-bench-metricsev-sim-results

6.3 Ablations and diagnostic examples

Read component removals and qualitative examples within their stated evaluation conditions.

Table 1. Embodied pretraining helps this fixed-position task, while the effect of adaptation depends on robot-state input. Original paper, p. 12 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. Use the retained table note to decode the two row switches: VidAW means GE-Base initialization and VidAda means task-specific video adaptation. Within each success metric, compare the with-state and without-state columns separately. Section 3.4 defines SR as the fraction of successful sub-steps, whereas E2E scores final completion and permits retries, so the two columns need not track each other. The experiment uses 305 demonstrations of moving a red cylinder into a paper cup at fixed positions, with 40,000 training steps per model. Comparing adjacent rows within one state column is more informative than treating the bold final row as a universal improvement. ev-ablationev-robot-eval

What it supports. With robot state, adding video adaptation to GE-Base initialization raises E2E from 0.81 to 0.89 and SR from 0.64 to 0.76: gains of 8 and 12 percentage points. The corresponding row without either intervention has E2E 0.15 and SR 0.05. This supports the value of embodied visual pretraining in the tested setup.

Where the evidence stops. Nearby prose calls the added intervention general video pretraining; the table defines task-specific adaptation. Without state, adaptation lowers E2E from 0.49 to 0.37, even though SR rises. Trial counts and uncertainty are absent, limiting causal strength and statistical interpretation.

7. Analysis & limitations

7.1 What the evidence leaves open

Source description

Authors identify a single-platform robotic pretraining corpus, parallel-jaw upper-body manipulation, and proxy evaluation with partial human validation. Dexterous hands, locomotion and broad heterogeneous-source robustness remain untested. ev-limitations

Reader analysis

Table 1 labels the added intervention as task-specific video adaptation, although nearby prose calls it general video pretraining. Its no-state E2E decrease rules out a universal adaptation benefit. Figures 8 and 11 provide no error bars or trial counts; human-ranking validation gives no correlation coefficient or annotator count. ev-ablationev-robot-evalev-transferev-human

Reader analysis

GE-Sim’s diagram uses a plus for motion/reference fusion, while caption and text specify concatenation; implementation remains unresolved. Action representation also needs clarification: the introduction calls GE-Act outputs torque trajectories, whereas GE-Sim explicitly accepts end-effector poses. ev-sim-symbolev-platformev-sim-interface

7.2 Questions for discussion

  1. Would Table 1’s adaptation gain persist with randomized positions and repeated seeds, particularly without robot state?
  2. How much do best-of-three selection and VLM judgment change video-model rankings?
  3. Can GE-Sim preserve policy rankings under contact errors and recovery attempts?

8. Reproducibility audit

8.1 Requirements and known gaps

Reader analysis

Reproduction requires pretrained video weights, calibrated synchronized views, instruction/action alignments and embodiment-specific control interfaces. Complete optimizer, learning-rate, batch-size and action-normalization recipes are absent. GE-Act reports 16 A100s for three-day pretraining and eight A100s for 12-hour video adaptation; final specialization is 36 hours in Section 3.2 but 24 hours in Figure 7. ev-base-trainingev-action-trainingev-training-discrepancyev-sim-interface

Reader analysis

Benchmark reproduction needs the adapted DINOv2 encoder, gripper detector, reference trajectories and VLM prompts. Preserve three-sample trajectory selection. Proposed checks should separately test the state/adaptation interaction and whether simulator rankings predict physical outcomes; neither follows automatically from video quality. ev-bench-dataev-bench-metricsev-ablationev-sim-resultsev-limitations

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: Replicate the state-by-adaptation interaction

Reader-proposed check, not an experiment performed here: start from the same GE-Base checkpoint and run the four combinations of robot state present/absent and task-specific video adaptation enabled/disabled. Keep the 305 demonstration trajectories, 40,000 action-training updates, camera inputs and evaluation retry budget identical. Repeat with multiple seeds and the same held-out initial configurations, recording both sub-step SR and E2E with trial counts and confidence intervals. Add randomized object positions as a separately reported stress test. A persistent no-state E2E decrease would support an interaction rather than a universal adaptation gain; disappearance across seeds would weaken that interpretation of Table 1. ev-ablationev-robot-evalev-action-training

Check 2: Test whether simulator scores preserve physical policy rankings

Reader-proposed check, not an experiment performed here: evaluate a fixed set of policy checkpoints on matched initial scenes in the real robot and both GE-Sim backbones. Include ordinary executions and controlled failed-grasp or pose-offset cases motivated by the simulator’s failure-data training. First replay identical recorded actions to separate rendering error from policy feedback; then allow closed-loop policies to act on generated observations. Keep the action horizon, camera calibration and retry budget fixed, and report every rollout alongside a separately labeled best-of-three metric. Compare physical E2E ordering with simulated ordering and track error growth over successive chunks. Good video scores with reversed policy rankings would falsify their use as a sufficient policy-evaluation surrogate. ev-sim-interfaceev-sim-trainingev-sim-loopev-sim-resultsev-bench-metricsev-robot-evalev-limitations

8.3 Reading coverage

Visual audit: PDF pages 1–23 were rendered and actually viewed, including the title/author/version page, all 18 figures, both tables, method equations, training hardware, evaluation definitions and limitations. All nine text chunks were read individually, including references through page 26; no appendix is present. Six original crops were inspected individually, with the narrow tables rendered at 400 DPI and the ablation crop rechecked after widening its margin. Figure 7’s update/duration disagreements, Figure 14’s fusion-symbol disagreement and Table 1’s prose/label disagreement are disclosed. References on pages 24–26 were read as text; external code, datasets, linked resources and separate supplements were outside this visual pass.

PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23. Appendix coverage: not present.

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 and abstract (p. 1)
  • 1 Introduction (pp. 1–4)
  • 2 GE-Base: World Foundation Model (pp. 4–7)
  • 2.1 Basic Architecture
  • 2.2 World Model Pre-training
  • 2.3 Robotic Manipulation Video Generation via GE-Base
  • 3 GE-Act: World Action Model (pp. 8–12)
  • 3.1 Basic Architecture
  • 3.2 Training Procedure
  • 3.3 Asynchronous Inference
  • 3.4 Action Planning via GE-Act on AgiBot G1
  • 4 Cross-Embodiment Generalization (pp. 12–15)
  • 4.1 Few-shot Adaptation
  • 4.2 Generalization to Agilex Cobot Magic
  • 4.3 Generalization to Dual Franka
  • 4.4 Generalization to RoboTwin
  • 5 GE-Sim: World Simulator (pp. 15–17)
  • 5.1 Hierarchical Action-conditioning Mechanism
  • 5.2 Training Procedure
  • 5.3 Action-conditioned Video Generation
  • 5.4 Closed-Loop Simulation
  • 6 EWMBench (pp. 18–22)
  • 6.1 Benchmark Dataset
  • 6.2 Evaluation Metrics
  • 6.3 World Model Evaluation
  • 6.4 Simulation Evaluation
  • 6.5 Metric-Human Consistency
  • 7 Related Works (pp. 22–23)
  • 8 Limitations (p. 23)
  • 9 Conclusion (p. 23)
  • Acknowledgment (pp. 23–24)
  • References (pp. 24–26)

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 extraction’s figure-image omission was addressed by separately inspecting PDF pages 1–23 and every final crop. References on pages 24–26 were read as text; there is no appendix in this PDF.
  • Version scope: the inspected title page identifies arXiv:2508.05635v3 [cs.RO], 4 November 2025. Its title and all 14 authors match the catalog. Earlier revisions and a separate ICLR 2026 proceedings edition were not supplied; differences from those editions and the catalog venue label cannot be verified.
  • Only the supplied primary PDF was reviewed. Project pages, code, model weights, datasets and separately linked benchmark papers were not inspected; no experiments were reproduced.

The visual audit above records the subsequent illustrated pass.

8.4 Traceable evidence

ev-identityPDF p. 1, title, author block, affiliations and arXiv margin stampInspect

The exact observed title and 14-author sequence match the catalog. The PDF is arXiv:2508.05635v3 [cs.RO], dated 4 November 2025; affiliations printed are AgiBot Genie Team, LV-NUS Lab and BUAA.

Go to primary source ↓
ev-platformPDF pp. 1–3, Abstract, Section 1 and Figure 1Inspect

The platform combines GE-Base, GE-Act, GE-Sim and EWMBench. Page 3 calls GE-Act outputs 54-step torque trajectories; the detailed action architecture later describes generic action tokens. Public release is promised, not established by this reading.

Go to primary source ↓
ev-basePDF pp. 4–5, Figure 3 and Section 2.1, unnumbered generation and conditioning equationsInspect

GE-Base conditions autoregressive chunks on initial observations, sparse history and language. It uses head and two wrist views, shared encoding, view/position embeddings, selected cross-view attention blocks and frozen T5-XXL. LTX-Video 2B and COSMOS2 2B are named. N denotes frames on p. 4 and perspectives on p. 5.

Go to primary source ↓
ev-base-trainingPDF pp. 5–6, Section 2.2, Data Curation, Stage I, Stage II and Figure 4Inspect

Approximately one million episodes total 2,967 hours. MR uses 57 frames at 3–30 Hz, four memory frames and eight latent frames; LF uses nine frames at 5 Hz, four memory frames and two latent frames. Encoder/decoder are frozen. Compute is 32 A100 GPUs for seven days plus three days. A complete optimization recipe is not given.

Go to primary source ↓
ev-memoryPDF p. 7, Section 2.3, inference paragraph and Figure 5Inspect

Video-generation inference samples memory frames uniformly at fixed intervals from earlier chunks. Figure 5 displays generated multi-view sequences rather than robot-execution measurements.

Go to primary source ↓
ev-actionPDF p. 8, Figure 6 and Sections 3–3.1, visual/action block equationsInspect

A 160M-parameter action decoder parallels LTX-based GE-Base, mirrors its depth with reduced width, and cross-attends to intermediate visual features. Noisy action tokens are refined by flow matching; explicit video generation is unnecessary for control.

Go to primary source ↓
ev-action-trainingPDF p. 9, Section 3.2, Pre-training and Task-specific adaptation tuningInspect

Only the action module updates in pretraining with GE-Base-LF fixed: four 5 Hz conditioning frames and 54-step 30 Hz targets, three days on 16 A100 GPUs. Video adaptation uses full AgiBot data and task data weighted 10×, taking 12 hours on eight A100s. Task action specialization updates backbone and action module on task-only data, stated as 36 hours on eight A100s.

Go to primary source ↓
ev-training-discrepancyPDF p. 9, Figure 7 graphic and caption versus Section 3.2Inspect

The graphic marks Action DiT trainable in stage 1, Video DiT in stage 2, and both in stage 3, matching the detailed procedure. The caption instead says the visual backbone is optimized first and the video encoder adapted later. Stage 3 is labeled 24 hours in the graphic but 36 hours in prose.

Go to primary source ↓
ev-fastPDF p. 10, Section 3.3, Asymmetric Denoising StrategyInspect

One visual denoising step supplies cached features for five action denoising steps. A 54-step forward pass takes a reported 200 ms on an onboard NVIDIA RTX 4090. Visual and action resolutions are stated as 5 Hz and 30 Hz; sensing-to-actuation latency statistics are absent.

Go to primary source ↓
ev-robot-evalPDF pp. 10–11, Figure 8, Section 3.4 and Evaluation ProtocolsInspect

Five G1 tasks compare slow/fast GE-Act and VLA baselines. SR measures completed sub-steps; E2E allows retries and scores final completion. The text specifies common protocols and demonstrations for discussed baselines. Figure 8 includes π₀ and lacks numerical bar annotations, error bars and evaluation counts.

Go to primary source ↓
ev-ablationPDF p. 12, Section 3.4 Analysis and Table 1, all rows and E2E/SR state columnsInspect

A fixed-position cylinder-to-cup task uses 305 demonstrations and 40,000 training steps. With state, GE-Base initialization yields E2E/SR 0.81/0.64, rising to 0.89/0.76 with VidAda. Without state those rows are 0.49/0.26 and 0.37/0.37. Neither initialization nor adaptation gives 0.15/0.05 with state. Caption defines VidAda as task-specific video adaptation; nearby prose attributes the best row to combining general video pretraining. Trial counts and uncertainty are absent.

Go to primary source ↓
ev-transferPDF pp. 12–14, Sections 4.1–4.3 and Figure 11(a,b) on p. 13Inspect

New embodiments need video adaptation and an action head trained from scratch; CLIP and video encoders remain frozen during video adaptation. Agilex cloth/box tasks and Dual Franka cloth folding each use 250 demonstrations, approximately one hour, with respective same adaptation data for GR00T N1, UniVLA and π₀. GE-Act has higher plotted SR/E2E; exact bar values, trial counts and uncertainty are unlabeled.

Go to primary source ↓
ev-robotwinPDF pp. 14–15, Section 4.4; PDF p. 13, Figure 11(c)Inspect

GE-Act jointly trains on four RoboTwin tasks using 50 demonstrations each; baselines use task-specific adaptation. Authors report advantages on three tasks and a deficit on lift pot, speculatively attributed to task interference.

Go to primary source ↓
ev-sim-interfacePDF p. 15, Sections 5 and 5.1Inspect

GE-Sim uses LTX-Video or COSMOS2 2B, a frozen CLIP reference-image anchor and camera-conditioned observations. Each arm has seven action coordinates: position, orientation and gripper openness; a K-step dual-arm trajectory is K×14.

Go to primary source ↓
ev-sim-conditioningPDF p. 16, Figure 14(a,b), Section 5.1, Pose2Image and Motion Vector Conditioning, Eqs. (1)–(2)Inspect

Calibrated projections encode position, orientation axes and gripper openness with arm-specific colors. Pose images and matching history frames share an encoder and are added in latent space. Position/orientation deltas are encoded, concatenated with the reference-image style token and injected through DiT cross-attention.

Go to primary source ↓
ev-sim-symbolPDF p. 16, Figure 14(a), caption, Eq. (1) and Motion Vector Conditioning paragraphInspect

The top merge of CLIP reference and motion features is a circled plus, but caption and text specify concatenation. The lower pose/history plus agrees with element-wise addition in Eq. (1). The two operations should not silently be treated as identical.

Go to primary source ↓
ev-sim-trainingPDF pp. 16–17, Section 5.2Inspect

GE-Sim initializes from GE-Base-MR and uses ground-truth trajectories on AgiBot-World-Beta, augmented by failed, incomplete and suboptimal executions from teleoperation and deployed robots. VAE and CLIP encoders freeze; other parameters use a video flow-matching loss. Failure-data count and full training configuration are unstated.

Go to primary source ↓
ev-sim-loopPDF p. 17, Section 5.4; PDF p. 16, Figure 14(b)Inspect

A policy proposes actions from images and instruction; GE-Sim generates the next video chunk for policy feedback. Reusing actions under different initial visual environments is proposed for data generation.

Go to primary source ↓
ev-bench-dataPDF p. 19, Section 6.1Inspect

Ten AgiBot-World-Beta test tasks are disjoint from million-episode pretraining tasks, with 100 videos per task and 4–10 annotated sub-actions. Trajectory selection uses voxelized end-effector paths and greedy low-overlap selection based on 3D IoU.

Go to primary source ↓
ev-bench-metricsPDF pp. 19–20, Section 6.2, Scene Consistency, Action Trajectories Quality and Motion Semantics MetricsInspect

Scene similarity uses adapted DINOv2 patch features. Generated gripper trajectories use inverse symmetric Hausdorff and NDTW distances; the best spatial match of three samples is retained. DYN uses amplitude-weighted inverse Wasserstein distances for velocity/acceleration. Semantic evaluation uses Qwen2.5-VL-7B-Instruct captions, BLEU, CLIP, logical-error penalties and CLIP diversity. ε is 10⁻⁸. Full prompts and detector training recipes are absent.

Go to primary source ↓
ev-bench-resultsPDF p. 20, Figure 17(a,b) and Section 6.3Inspect

With text-and-head-image conditioning, GE-Base scene/motion/semantics/score are 0.9427/1.6676/2.0907/4.7010; Kling is 0.8888/0.9440/2.0370/3.8698; LTX total is 2.9676. Six models are shown although Section 6.3 says seven. GE-Base uses robot-adapted LTX.

Go to primary source ↓
ev-sim-resultsPDF p. 21, Section 6.4 and Table 2, both model rows and both metric blocksInspect

Ground-truth-action-conditioned GE-Sim: LTX BLEU/CLIP/DYN/diversity/PSNR/SA/logic/TA/scene are 0.33/90.8/0.78/0.011/19.9/0.94/0.97/0.98/0.90. COSMOS2, labeled COSMOS, has 0.31/90.2/0.85/0.010/20.7/0.87/0.97/0.97/0.91. These are video metrics, not physical success.

Go to primary source ↓
ev-humanPDF pp. 21–22, Figure 18 and Section 6.5Inspect

Four video models are compared using human rankings, EWMBench and VBench. Authors describe aggregated ordinal judgments and repeated review. The figure supports concordant EWMBench ordering; annotator count, uncertainty and numerical correlation are not reported.

Go to primary source ↓
ev-limitationsPDF p. 23, Section 8, all three limitation bulletsInspect

Authors identify single-platform robotic pretraining, parallel-jaw upper-body manipulation and proxy evaluation with partial human validation. Heterogeneous-source robustness, dexterous hands, locomotion and reliable automated task-success evaluation remain open.

Go to primary source ↓

8.5 Primary sources

Scroll across the image to inspect details. Press Esc to close.