PaperlayerView exact revision on arXiv

Source-linked AI summary

Ropedia W0: Scaling Loco-Manipulation World Models across Embodiments with Human Experience

Tianqi Liu, Hanfei Zhao, Qifeng Liang, Hui En Pang, Jiarui Xu, Sayang Mu, Penghao Qiu, YingSheng Zhang, Jun Wang, Ziang Cao, Mingze Sun, Yukang Cao, Fangzhou Hong, Zhaoxi Chen, Ziwei Liu

arXiv:2610.26855v1cs.CVcs.RO

TL;DR

World models have yet to jointly address locomotion, manipulation, and generalization across diverse embodiments, with limited multimodal data and supervision. Ropedia W0 unifies human and robot experience through a shared kinematic representation; it reports strong cross-embodiment prediction and supports real-time rollouts and embodied decision-making applications.

  • Problem

    Existing models and datasets leave joint locomotion, manipulation, and cross-embodiment world modeling underexplored, with incomplete or unreliable multimodal annotations.

    Source: §1 · p. 2, §3 · p. 4

  • Method

    Ropedia W0 trains on curated human and robot videos using shared camera-pose and projected-skeleton conditioning, with Xp-Data and the Xp-Teleop benchmark.

    Source: §1 · pp. 2–3

  • Results

    Ropedia W0 achieves state-of-the-art cross-embodiment prediction, including on Xp-Teleop, with strong zero-shot generalization and data-efficient adaptation across diverse benchmarks.

    Source: §1 · p. 3, §7 · p. 18

  • Takeaways & Limitations

    Distillation supports real-time, minute-long rollouts and live VR interaction, while experiments demonstrate policy evaluation, planning, and world-action modeling applications.

    Source: §1 · p. 3

    The kinematic interface captures camera and manipulator motion but not physical contact or object dynamics.

    Source: F. Limitations and Future Work · p. 54

Abstract

from arXiv · show

World models enable embodied agents to anticipate action consequences, yet jointly modeling locomotion and manipulation across diverse embodiments remains challenging due to heterogeneous action spaces and limited multimodal supervision. We present Ropedia W0, the first egocentric loco-manipulation world foundation model spanning humans and robots with diverse embodiments. We curate Xp-Data, comprising 10k hours of egocentric human data and 3k hours of robot data with unified multimodal annotations and systematic quality control. A unified kinematic representation encodes locomotion through camera poses and manipulation through projected skeletons, with robot skeletons combining embodiment-specific forward kinematics with extensive manual keypoint annotation and correction. We also introduce Xp-Teleop, a long-horizon loco-manipulation benchmark based on high-quality dexterous humanoid teleoperation data. Evaluations demonstrate state-of-the-art performance across five cross-embodiment benchmarks (DreamDojo, $\Psi_0$, $\omega$-HOME, Xp-Teleop, and AgiBot) and on held-out human egocentric video, from zero-shot inference to few-shot and full-data post-training. Scaling human pre-training data improves performance on both human and robot evaluations. After distillation, our model supports real-time, minute-long rollouts, enabling live VR interaction. Applications in policy evaluation, model-based planning, and world-action modeling further demonstrate the model's utility for embodied decision-making.

1. Introduction

Ropedia W0 addresses the challenge of jointly modeling locomotion and manipulation across embodiments, where existing data and models leave important gaps. It combines human and robot data through a shared kinematic representation and reports strong cross-embodiment results and downstream applications.

1. Introduction

  • Existing world models largely focus on fixed-base gripper manipulation, leaving locomotion-driven viewpoint changes, dexterous interactions, and cross-embodiment generalization underexplored.Some cross-embodiment approaches require dedicated control interfaces or additional adaptation for each new embodiment.Source: §1 · p. 2
  • Public egocentric datasets often lack reliable metric camera trajectories and hand poses, use heterogeneous formats, and remain limited in scale.Robot demonstrations across tasks and embodiments are expensive to collect, while human video offers complementary navigation and interaction experience.Source: §1 · p. 2
  • A shared explicit kinematic representation combines metric camera poses for locomotion with projected skeletons for manipulation across human and robot training stages.Robot skeletons combine embodiment-specific forward kinematics with manually annotated or corrected keypoints in an image-aligned format.Source: §1 · p. 3
  • Xp-Data combines 10k hours of human data and 3k hours of robot data with unified geometric, kinematic, and semantic annotations and quality checks.Robot keypoints receive manual annotation and correction.Source: §1 · p. 3
  • Ropedia W0 achieves state-of-the-art cross-embodiment prediction, including on Xp-Teleop, and supports real-time interaction, policy evaluation, planning, and world-action modeling.Scaling human pre-training data improves prediction on human videos and transfer to robots; distillation enables minute-long rollouts and live VR interaction.Source: §1 · p. 3

2. Related Work

Prior work explores video world models, robot-specific action conditioning, and human-to-robot transfer, but sustained cross-embodiment loco-manipulation remains a distinct challenge. Ropedia W0 uses an explicit kinematic interface and supports evaluation, planning, and action learning applications.

Robotic video world models

  • Robotic world models use embodiment-specific action encodings, inferred latent actions, or spatially aligned renderings; cross-embodiment methods also use action heads, projections, or adaptation.Latent actions can make unlabeled footage usable but may entangle actions with other visual changes.Source: §2 · pp. 3–4
  • AnyWorld recomposes human interactions into robot experiences but conditions on source-human skeletons and is evaluated mainly on manipulation, leaving sustained loco-manipulation untested [22].Ropedia W0 instead uses an explicit kinematic interface across humans, gripper robots, and dexterous humanoids.Source: §2 · p. 4

Human and robot datasets

  • Existing human and robot datasets provide varied activity, pose, and action coverage, but Xp-Data unifies annotations and Xp-Teleop targets long-horizon dexterous humanoid loco-manipulation.Prior humanoid datasets include locomotion or dexterous hands, while evaluation datasets cover held-out clips and multimodal world-model benchmarks.Source: §2 · p. 4

Policy evaluation, planning, and learning

  • World models support policy evaluation and planning through predicted outcomes, while Ropedia W0 transfers its pretrained video backbone to initialize a world-action model without a dedicated action head.Vision-language-action models map observations and language to actions, increasingly using human video for training.Source: §2 · p. 4

3. Data

Existing datasets often omit camera geometry and have noisy hand annotations, limiting their suitability for general-purpose world models. Xp-Data addresses these gaps with curated human and robot recordings, unified multimodal annotations, and quality control.

3. Data

  • Existing datasets often lack camera poses and contain noisy or inaccurate hand annotations, leaving them insufficient for general-purpose world-model training.Cross-embodiment loco-manipulation requires shared annotations for viewpoint changes and manipulator motion.Source: §3 · p. 4
  • Xp-Data contains over 10k human and 3k robot training hours, with synchronized RGB, camera trajectories, manipulator skeletons, and captions in a shared schema.It combines public datasets with over 6,000 hours of in-house four-fisheye-camera video; wrist poses and wide-FOV skeletons are included where available.Source: §3 · pp. 4–5
  • Xp-Teleop provides about one held-out hour of long-horizon dexterous humanoid loco-manipulation data recorded in the same schema.Its episodes combine locomotion and manipulation, addressing a benchmark scope distinct from tabletop teleoperation.Source: §3 · p. 4
  • The curation pipeline filters low-quality or irrelevant video, annotates geometry, kinematics, and hierarchical semantics, then checks annotation completeness, correctness, and cross-modal consistency.Robot keypoints combine URDF forward kinematics with manual annotation and correction; caption checks pair automated verification with human review.Source: Figure 2 · p. 5, §3 · p. 6
  • Quality-checked subtasks become standardized 81-frame, 640 × 480, 16 fps clips with synchronized multimodal signals in a shared human–robot format.The clips include RGB, projected skeletons, camera parameters, and captions, plus wide-FOV skeletons and wrist poses where available.Source: Figure 2 · p. 5, §3 · p. 6
  • Robot skeletons use forward-kinematics estimates where available, supplemented by manual image-space keypoint annotation and correction when projections misalign.Wrist poses are derived through forward kinematics where available.Source: §3 · p. 6

4. Model

Ropedia W0 unifies locomotion and manipulation as camera poses and projected skeletons, then conditions a video diffusion model on those controls. Human pre-training and human–robot mid-training support cross-embodiment prediction, while distillation enables minute-long streaming rollouts.

4.1. Representation: Unified Kinematic Control

  • Camera poses encode locomotion-induced viewpoint changes as dense Plücker-ray maps, while projected skeletons encode manipulation across embodiments.Camera translations use metric units; skeletons preserve gripper openings and dexterous finger configurations under camera and manipulator motion.Source: §4.1 · pp. 7–8
  • Robot skeletons combine embodiment-native forward kinematics with manual image-space keypoint annotation and correction.Where joint states and calibration are available, forward kinematics provides initial keypoint projections.Source: §4.1 · p. 8

4.2. Architecture: Kinematics-Conditioned World Model

  • A latent video diffusion transformer predicts future frames from the initial observation, text, and kinematic controls, injecting aligned conditions through projections and additional context through cross-attention.Camera and skeleton latents are aligned with video latents in space and time; optional wrist-pose and wide-FOV conditions have separate injection paths.Source: §4.2 · p. 9

4.3. Training: From Human Experience to Cross-Embodiment

  • Training proceeds from human egocentric pre-training to balanced human–robot mid-training using the shared conditioning interface.Mid-training samples human and robot data with equal probability and adds robot-specific appearance, geometry, and interaction patterns while retaining human data exposure.Source: §4.3 · p. 9

4.4. Distillation: Real-Time Interactive Rollouts

  • 21.3 fps at 640 × 480 enables minute-long interactive rollouts from the distilled student.The model uses a 4-step generator adapted for long-video streaming with a bounded KV cache.Source: §4.4 · pp. 9–10

5. Xp-Teleop

Xp-Teleop is a humanoid teleoperation benchmark for long-horizon prediction during locomotion and dexterous manipulation. Its synchronized recordings combine those behaviors within everyday household tasks.

5. Xp-Teleop

  • 27 long-horizon takes provide about one hour of egocentric video at 30 fps, recorded on a Unitree G1 with two six-DoF dexterous hands.Two operators control locomotion and manipulation concurrently; the benchmark spans six household activity groups.Source: §5 · p. 10, Figure 5 · p. 11
  • Each take uses synchronized RGB-D, camera poses, projected skeletons, robot states, language, tactile, and head-IMU signals.The recordings use the same schema as Xp-Data and include 543 human-annotated atomic segments across 30 skill classes.Source: §5 · p. 10, Figure 5 · p. 11
  • Keeping failed grasps and recoveries lets the benchmark probe simulation of off-nominal outcomes, not only successful executions.Source: §5 · p. 10
  • Table 1 compares Xp-Teleop with other humanoid teleoperation datasets by task coverage and available modalities.Source: §5 · p. 10

6. Experiments

Ropedia W0 is evaluated across five cross-embodiment benchmarks and held-out human video, with analyses of scaling, conditioning, and downstream uses. Results include strong transfer and real-time rollouts after distillation.

6.1. Setup

  • The evaluation covers five cross-embodiment benchmarks, held-out human video, three post-training settings, and both human-only pretraining and human-plus-robot midtraining.Zero-shot means direct evaluation on a target dataset never used for training; the target embodiment may appear in other midtraining datasets.Source: §6.1 · pp. 11–12

6.2. Benchmark Results

  • A14B midtrain leads all three adaptation settings on Xp-Teleop, and its zero-shot result exceeds every baseline’s full-data result.It also leads on long-horizon autoregressive rollouts, at approximately 32 seconds on ω-HOME and 30 seconds on Xp-Teleop.Source: §6.2 · p. 12, Table 4 · p. 14

6.3. Ablation Studies and Analysis

  • Removing skeleton conditioning reduces PSNR by 3.8 dB for A14B and 1.8 dB for 5B; removing camera conditioning causes smaller average drops, but a 1.3 dB drop on moving-viewpoint embodiments.The findings support complementary roles for both conditions, with camera conditioning especially important for loco-manipulation.Source: §6.3 · pp. 13–14
  • With 1% of post-training data, Ropedia W0 5B improves PSNR by 4.8/7.1/6.6 dB over the official base and outperforms that base post-trained on 100% of the data.Its advantage remains approximately 2.3 dB under full-data post-training; increasing human-video pre-training also improves held-out human prediction and transfer across robot benchmarks.Source: §6.3 · pp. 14–15, Figure 6 · p. 15, Figure 7 · p. 16
  • Scaling from 5B to A14B improves PSNR across the reported regimes, while the distilled 4-step 5B student reaches 21.3 fps and matches teacher fidelity through minute-long rollouts.The student runs at 7× the 50-step teacher’s speed, exceeding the stated 20 fps real-time line.Source: §6.3 · p. 15

6.4. Downstream Applications

  • As a proxy environment, Ropedia W0’s task-progress scores closely match ground truth and preserve policy rankings; closed-loop planning reaches 92% progress in RoboTwin 2.0 and 55% on the real G1.These planning results compare with random selection / the best fixed checkpoint at 53% / 85% in simulation and 26% / 45% on the real G1.Source: §6.4 · pp. 16–17, Figure 9 · p. 17
  • Initializing FastWAM’s video backbone from Ropedia W0 improves LIBERO-Plus success from 50.06% to 56.82%; live teleoperation uses a distilled student to generate egocentric video in real time across embodiments.On a real G1 loco-manipulation task, the initialization reaches 56.7% task progress versus 33.3% for FastWAM and 50.0% for the Ψ0 VLA policy.Source: §6.4 · p. 17, Table 6 · p. 17, Figure 10 · p. 18

7. Conclusion

Ropedia W0 is a unified egocentric world model for loco-manipulation across humans and robots, using camera poses and projected skeletons in an explicit kinematic representation. The paper concludes that human-video scaling improves human prediction and robot transfer, while the model also supports downstream decision-making and interactive simulation.

7. Conclusion

  • The conclusion presents Ropedia W0 as a unified world model for loco-manipulation across humans and robots with grippers or dexterous hands.Its shared representation combines camera poses for viewpoint motion with projected skeletons for manipulation, preserving embodiment-native geometry in a shared visual format.Source: §7 · p. 18
  • The paper reports strong zero-shot generalization, data-efficient adaptation, and competitive prediction across diverse benchmarks, including Xp-Teleop.Increasing human-video pre-training improves held-out human-video prediction and transfer to robot benchmarks.Source: §7 · p. 18
  • The conclusion identifies policy evaluation, model-based planning, and world–action modeling as uses beyond video prediction, with distillation supporting real-time interactive simulation.Source: §7 · p. 18

A. Data and Benchmark Details

  • The appendix documents the curated training corpus and unified annotations, then describes Xp-Teleop’s collection setup and dataset-comparison criteria.Source: A. Data and Benchmark Details · p. 28

A.1. Data Curation and Annotation Examples

Xp-Data combines human and robot recordings with aligned motion annotations, while Xp-Teleop supplies long-horizon humanoid teleoperation data. The section illustrates the corpus coverage and defines how dataset modalities and task horizons are characterized.

A.1. Data Curation and Annotation Examples

  • Xp-Data’s curation funnel screens, annotates, and quality-checks recordings, yielding approximately 3,000 robot hours and 10,000 human hours.Figures S2–S4 show projected hand or manipulator skeletons and per-window camera trajectories across the human and robot examples.Source: Figure S1 · p. 28, §A.1 · p. 28, Figure S2 · p. 29

A.2. Xp-Teleop Collection and Benchmark Details

  • Xp-Teleop retains all operator-initiated takes, including failed grasps and recoveries, to evaluate simulation of both successful and off-nominal outcomes.Two operators concurrently control locomotion and manipulation; synchronized recordings yield camera poses and image-projected hand skeletons.Source: §A.2 · p. 32
  • Xp-Teleop contains 27 teleoperation takes totaling about one hour, divided into 543 annotated segments across 30 skill classes.Each take contains 8–42 segments, with a median of 18; captions identify objects and task progression.Source: §A.2 · p. 32
  • The dataset comparison separates tabletop teleoperation, where legs remain locked, from loco-manipulation, where locomotion can occur alongside manipulation.Task-horizon columns report episode-level means; duration is computed from recorded frames and frame rate.Source: §A.2 · p. 32
  • Subtask counts measure sequential instruction sub-goals by splitting instructions into clauses and counting clauses beginning with listed action verbs.Locomotion and manipulation clauses count equally, so this is an annotation-level task-horizon measure, not a platform-independent action count.Source: §A.2 · p. 33
  • Dataset modality criteria distinguish usable depth and explicit IMU or tactile data from estimated poses, and count whole-body joints only when recorded states include legs.Image-space keypoints may be provided or reconstructed from released joint states, camera calibration, and a public URDF.Source: §A.2 · pp. 33–34

B. Model and Implementation Details

The implementation pairs pretrained video-diffusion backbones with camera and skeleton conditioning, then distills generation for faster autoregressive and long-video streaming. An optional wide-field-of-view skeleton branch addresses manipulators cropped by head cameras.

C. Evaluation Protocols

B.1. Backbone and Conditioning

  • Ropedia W0 uses Wan2.2 pretrained video-diffusion backbones at 5B dense and approximately 14B active-parameter MoE scales.Training retains the first temporal latent as clean visual context, applies noise and loss only to future latents, and adds camera-ray and skeleton branches through zero-initialized projections.Source: §B.1 · p. 34

B.2. Distillation and Interactive Simulation

  • Causal Forcing initializes a 4-step autoregressive generator, asymmetric DMD trains it against real- and fake-score models, and LongLive extends it to streaming.LongLive uses within-block bidirectional and cross-block causal attention with a bounded KV cache.Source: §B.2 · p. 35, Figure S5 · p. 35, Figure S6 · p. 36
  • At 640 × 480, the 4-step student generates 125 frames in about 6 seconds including VAE decoding (≈21 fps), a 7× end-to-end speed-up over the 50-step teacher.It streams its first block in roughly 1.4 seconds; Table S1 reports that the 4-step student matches or exceeds the teacher at every rollout length.Source: §B.2 · p. 36, Table S1 · p. 36

B.3. Wide-FOV Skeleton Conditioning

  • The optional wide-FOV branch supplies arm configuration missing from the narrow head-camera view; on WorldArena, 1.5× rendering raises in-frame manipulator coverage from 0.37 to 0.56.In the two examples, border-band keypoint error falls from 8.9 to 6.9 px and from 14.0 to 11.7 px, respectively.Source: §B.3 · p. 37, Figure S7 · p. 37, Figure S8 · p. 38

C.1. Human Egocentric Video Evaluation

The human-video evaluation uses aligned initial conditions and defined camera, hand, and video-quality metrics. The scaling protocol distinguishes zero-shot human evaluation from target-benchmark post-training, while supplementary tables break out dataset-specific results.

C.1. Human Egocentric Video Evaluation

  • Human evaluation uses 200 clips from EgoSuite and OpenAoE, sampled across high, medium, and low camera-motion strata, with 81 frames at 640 × 480 and 16 fps.Methods receive the same initial frame and, when supported, caption; baselines retain their native conditioning interfaces.Source: §C.1 · p. 39
  • ATE, RPE, and RRE measure aligned trajectory, stepwise translation, and relative rotation errors; hand metrics distinguish missed-detection-penalized PCK@.1 from matched-hand PCKm@.1 and detection rate.Generated-video camera poses are estimated with DA3, and hand detections use MediaPipe at confidence 0.3.Source: §C.1 · p. 39

C.2. Scaling Protocol

  • Scaling checkpoints represent unique human-video hours processed with a fixed data mixture and a single epoch; human-video evaluation is zero-shot, while transfer and DreamDojo evaluations use target post-training.Transfer post-training uses one clip for 50 steps on Ψ0 and 144 clips for five epochs on Xp-Teleop.Source: §C.2 · p. 39

D. Additional Quantitative Results

  • The supplementary section provides per-dataset human results and additional seen/unseen robot breakdowns beyond the main benchmark summaries.The main summaries are in Tables 4, 5, and 2.Source: D. Additional Quantitative Results · p. 40

D.1. Human Evaluation by Dataset; D.2. Seen and Unseen Robot Evaluation

  • Table S2 separates the pooled human-video results by EgoSuite and OpenAoE, while Table S3 breaks dexterous-humanoid results into seen and unseen subsets under three adaptation settings.Table S2 reports ATE in centimeters and PSNR in dB; Table S3 reports PSNR and the seen–unseen gap.Source: §D.1 · p. 40, Table S2 · p. 40, §D.2 · p. 40

D.3. Long-Horizon Rollout Evaluation

Supplementary evaluations examine long rollouts, conditioning and pretraining ablations, and comparisons across WorldArena and robot and household clips. Ropedia W0 leads BWM on most reported WorldArena metrics, with the margin chiefly associated with trajectory accuracy.

D.3. Long-Horizon Rollout Evaluation

  • Tables S4 and S5 report approximately 32-second zero-shot ω-HOME and 30-second Xp-Teleop rollouts, including aggregate quality and changes over rollout time.The tables describe 12 chained windows for ω-HOME and 11 autoregressively chained windows for Xp-Teleop.Source: §D.3 · p. 41, Table S4 · p. 41, Table S5 · p. 42

D.4. Conditioning and Pretraining Ablations

  • Table S6 evaluates camera and skeleton conditioning across 400 windows and eight embodiments, while Table S7 compares pretrained and public-weight initializations at matched post-training budgets.The Cosmos3 comparison uses the same robot mixture as Ropedia W0 mid-training; neither model is post-trained on evaluation windows.Source: Table S6 · p. 42, Table S7 · p. 42, §D.4 · p. 42

D.5. WorldArena Comparison against BWM

  • Ropedia W0 leads BWM + SR on 12 of 15 WorldArena metrics and by +5.9 EWMScore, with the margin dominated by +29.1 Trajectory Accuracy.The comparison uses WorldArena Track 1’s 1,000-episode fixed_scene_task test set and the same official pipeline and inputs.Source: §D.5 · p. 43, Table S8 · p. 43
  • The WorldArena qualitative panels compare aligned ground-truth, Ropedia W0, and BWM frames for drawer-and-box and cup-placement tasks.Ropedia W0 follows the commanded arm trajectory and completes both tasks, while BWM’s arm drifts and does not open the drawer.Source: Figure S9 · p. 44

E. Qualitative Comparisons

  • Qualitative comparisons align methods on the same initial frame and conditioning through their native interfaces, with predicted and ground-truth frames aligned in time.Project-page videos use the same displayed frame indices across methods.Source: E. Qualitative Comparisons · p. 44

E.1. Cross-Embodiment Robot Prediction

  • Held-out robot comparisons cover Fourier, AgiBot, Baihu, Dexora, RH20T, and DROID clips, spanning different manipulators, objects, and interaction types.The figures show input skeleton overlays and identify the adaptation setting.Source: §E.1 · p. 44, Figure S10 · p. 45, Figure S11 · p. 46

E.2. Zero-Shot Loco-Manipulation

  • Zero-shot ω-HOME comparisons use household tasks not seen in training and show generated hand motion, object interactions, and scene appearance against aligned ground-truth frames.Examples include reaching into a fridge, pushing a chair, and sorting or placing fruit.Source: §E.2 · pp. 48–49, Figure S13 · p. 48, Figure S14 · p. 49

E.3. Post-Training and Data Efficiency

Post-training examples show Ropedia W0 tracking commanded hands and matching ground truth, while limited-data initialization preserves detail lost by public video-generation weights. Additional examples cover multiple post-training budgets and 32-second rollouts.

E.3. Post-Training and Data Efficiency

  • On the ω-HOME fruit-bucket task, Ropedia W0 follows commanded hands and matches ground truth zero-shot and after post-training, while baselines lose the arm or drift.Source: Figure S15 · p. 50, §E.3 · p. 50
  • With only 1% of GR-1 data, pretrained Ropedia W0 yields sharper predictions and preserves object and hand detail better than public video-generation initialization.This comparison covers both In-lab and EgoDex settings.Source: §E.3 · p. 50, Figure S16 · p. 51, Figure S17 · p. 51
  • ω-HOME examples compare zero-shot, few-shot, and one-epoch post-training predictions for table wiping and towel placement in a washing machine.The figure uses the same frame index across settings and overlays the input skeleton conditions.Source: Figure S18 · p. 52, §E.3 · p. 52

E.4. Additional Long-Horizon Rollouts

  • Two additional generated sequences show 32-second rollouts for collecting fruit from a closet and picking up bok choy, with supplied camera and hand conditions.Frames are sampled every 6.4 seconds; quantitative long-horizon comparisons are reported elsewhere.Source: §E.4 · p. 53, Figure S19 · p. 53

F. Limitations and Future Work

The authors identify limits in rollout horizon, scaling scope, and physical modeling, and describe directions for extending each. The kinematic interface represents camera and manipulator motion but not contact or object dynamics.

F. Limitations and Future Work

  • The distilled student supports minute-long rollouts, while stronger memory or periodic re-grounding are proposed for stable interactive simulation at longer horizons.Source: F. Limitations and Future Work · p. 54
  • Scaling studies reach up to a 14B-active model and vary human-video data while holding the data mixture fixed; broader model, embodiment, and task scaling remain future directions.Source: F. Limitations and Future Work · p. 54
  • The kinematic interface captures camera and manipulator motion, not physical contact or object dynamics; physically grounded 3D assets and scenes are proposed to model underlying dynamics.Source: F. Limitations and Future Work · p. 54

Related papers

Loading 2610.26855v1…