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

cs.CVOct 19, 2026

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

Authors: Tianqi Liu, Hanfei Zhao, Qifeng Liang, Hui En Pang, Jiarui Xu, Sayang Mu, Penghao Qiu, YingSheng Zhang,

Organizations: Nanyang Technological University · Ropedia · Duke University · University of California, San Diego ·

Abstract

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, $Ψ_0$ , $ω$ -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.

Figures & tables

Figure 2: Xp-Data curation pipeline. Heterogeneous human and robot recordings pass through five stages. (1) Data sources : 14 public human egocentric datasets, 6,000 h of in-house four-camera recordings, and 13 public robot datasets. (2) Video filtering removes low-quality or task-irrelevant clips with a vision-language model. (3) Multimodal annotation produces geometry (metric camera poses via SLAM and DA3 [ 69 ] ), kinematics (human hands via WiLoR [ 91 ] and MANO [ 96 ] , robot keypoints via URDF forward kinematics, manual annotation, and correction), and semantics (hierarchical task, subtask, and atomic-action captions). (4) Quality check verifies completeness, correctness, and cross-modal consistency, pairing automated checks with a caption cascade (Gemini [ 25 ] annotation, Qwen [ 6 ] verification, and human review). (5) Training-clip construction cuts quality-verified subtasks into standardized 81-frame, $640\times 480$ , 16 fps multimodal clips, yielding the unified cross-embodiment corpus.

Figure 3: Overview of Xp-Data. Top: caption terms. Middle: environment and task distributions for robot (inner ring) and human (outer ring) data. Bottom: unified skeleton annotations across human hands, dexterous robot hands, and parallel grippers.

Figure 4: Overview of Ropedia W0. Left: Human pose annotations and robot forward kinematics provide kinematic structure and wrist poses; manual robot-keypoint annotation and correction align robot skeletons with the observed images. Middle: Camera trajectories are encoded as Plücker-ray maps, while manipulation is represented by skeleton maps, optionally augmented with wrist-pose and wide-FOV renderings. Right: A video diffusion transformer predicts future frames from an initial image, a text prompt, and these control signals. Pixel-aligned conditions are fused through additive projections, while wide-FOV features are injected via cross-attention.

Dataset

Robot / hand

Capabilities

Task horizon

Modalities

Loco.

Dex.

Duration (s)

Loco. range (m)

Subtasks

RGB@fps

Dpt.

IMU

WB joints

Tac.

Cal.

Kpts.

Tabletop teleoperation

RoboMIND (Tien Kung)

Tien Kung / Inspire

\times

\checkmark

21.7

2.2

640{\times}480

@30

\checkmark

\times

\times

\times

\times

\times

GR00T-Teleop-GR1

GR-1 / Fourier

\times

\checkmark

14.3

0.21

1.0

1280{\times}800

@20

\times

\times

\checkmark

\times

\times

\times

Unitree G1 Dex3

G1 / Dex3

\times

\checkmark

27.4

1.0

640{\times}480

@30

\times

\times

\times

\times

\times

\times

Unitree G1 Dex1

G1 / Dex1 gripper

\times

\times

28.7

1.1

640{\times}480

@30

\times

\times

\checkmark

\times

\times

\times

Table 1: Comparison of humanoid teleoperation datasets across capabilities (locomotion and dexterity), task horizon (duration, locomotion range, and number of subtasks), and available recording modalities (RGB frame rate, depth, IMU, whole-body joints, tactile, calibration, and keypoints). A check vs. cross marks whether each property is provided; datasets are grouped by whether the lower body is used during collection (tabletop vs. whole-body). Xp-Teleop is the only one offering synchronized long-horizon loco-manipulation with the full modality set.

Figure 5: Overview of Xp-Teleop. About one hour of dexterous-humanoid teleoperation spanning six everyday household activity groups, from tidying and food preparation to laundry and cleaning (a). Each take is a long-horizon episode that chains locomotion with manipulation, such as putting objects into a box or folding a jacket on the bed (b). Takes are recorded as a single synchronized multimodal stream (egocentric RGB-D, 6-DoF camera pose, projected hand/robot skeletons and masks, 29-joint robot state, time-aligned language segments, per-hand tactile, and head IMU; c), and their 543 annotated segments are dominated by loco-manipulation verbs such as reach , walk , and grasp (d).

Method	(a) Camera and hand action following
Method	ATE $\downarrow$	RPE $\downarrow$	RRE $\downarrow$	HandL2 $\downarrow$	[email protected] $\uparrow$	[email protected] $\uparrow$	Det $\uparrow$
Cosmos3-Nano (16B)	1.931	0.170	1.584	46.9	78.94	84.03	92.08
Cosmos3-Super (64B)	2.031	0.178	2.045	48.3	77.89	83.32	91.83
OSCAR (2B)	3.480	0.273	6.258	42.1	77.60	84.46	90.55
RynnWorld-Teleop (5B)	3.543	0.274	6.109	49.1	74.67	82.37	88.78
DreamDojo (2B)	2.652	0.224	4.716	69.8	58.84	70.14	81.44

Table 2: Results on a 200-clip EgoSuite [ 68 ] and OpenAoE [ 66 ] evaluation set (100 clips each). Zero-shot egocentric human-video prediction, reported as (a) camera- and hand-action following (trajectory errors ATE/RPE/RRE, hand position error HandL2, and PCK-based hand-keypoint accuracy with detection rate Det) and (b) video prediction and generation quality (PSNR/SSIM/LPIPS, VBench [ 51 ] consistency/dynamics/aesthetics, and FVD [ 106 ] ). ATE/RPE are in cm, RRE in degrees, HandL2 in pixels, and PSNR in dB; PCK/PCKm/Det and Cons./Dyn/Aes are reported as percentages. Best bold, second-best underlined; full protocol in Sec. C.1 .

Data	Model	In-lab			EgoDex			DreamDojo-HV
Data	Model	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
1%	DreamDojo (2B)	18.840	0.731	0.254	18.584	0.749	0.243	16.996	0.724	0.256
	DreamDojo (14B)	18.983	0.743	0.244	18.070	0.747	0.246	16.643	0.716	0.262
	Ropedia W0 (5B)	20.854	0.764	0.213	20.569	0.778	0.208	19.045	0.760	0.217
	Ropedia W0 (A14B)	23.028	0.826	0.190	22.354	0.830	0.188	20.788	0.810	0.198
10%	DreamDojo (2B)	19.238	0.744	0.244	19.091	0.768	0.230	17.210	0.729	0.250

Table 3: Comparison on the DreamDojo [ 34 ] benchmarks under different post-training settings. Despite having only 5B parameters, our model consistently outperforms the larger DreamDojo (14B) baseline across all metrics. Best results are in bold , and second best are underlined .

Benchmark	Model / initialization	No post-training			Few-shot			Full-data
Benchmark	Model / initialization	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
$\Psi_{0}$	DreamDojo (14B)	14.45	0.580	0.432	14.43	0.579	0.438	13.78	0.569	0.449
	Cosmos3-Nano (16B)	13.74	0.563	0.465	14.71	0.584	0.409	16.55	0.636	0.320
	Cosmos3-Super (64B)	13.78	0.562	0.461	14.94	0.591	0.400	15.73	0.613	0.352
	Ropedia W0 (5B), pretrain	15.17	0.587	0.394	15.59	0.606	0.364	16.26	0.625	0.330
	Ropedia W0 (5B), midtrain	15.22	0.594	0.385	15.59	0.603	0.369	16.41	0.628	0.325

Table 4: World-model evaluation on

\Psi_{0}

[ 117 ] ,

\omega

-HOME [ 65 ] , and Xp-Teleop under three post-training settings: no post-training (zero-shot transfer), few-shot, and full-data, reporting PSNR, SSIM, and LPIPS. For Ropedia W0 we report both the pretrain (human-only) and midtrain (human + robot) stages at the 5B and A14B scales. Best bold, second-best underlined.

Model / initialization	Fidelity			Video quality
Model / initialization	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	S-Cons. $\uparrow$	B-Cons. $\uparrow$	Aes. $\uparrow$	Dyn.
DreamDojo (2B), AgiBot post-trained	18.26	0.750	0.218	0.9383	0.9520	0.4318	0.54
DreamDojo (14B), AgiBot post-trained	18.30	0.755	0.210	0.9416	0.9526	0.4374	0.48
Cosmos3-Nano (16B)	17.70	0.742	0.230	0.9159	0.9262	0.4373	0.72
Cosmos3-Super (64B)	17.28	0.732	0.241	0.9181	0.9274	0.4396	0.68
OSCAR (2B)	18.56	0.746	0.217	0.9281	0.9458	0.4477	0.66

Table 5: Comparison on AgiBot [ 15 ] (50 held-out gripper windows). VBench [ 51 ] consistency, aesthetics, and dynamic degree. Best bold, second-best underlined.

Figure 6: Scaling on held-out human video. Zero-shot evaluation of the 5B model on 200 held-out egocentric clips versus human-video pre-training hours (log scale). The eight panels report camera ATE, RPE, and RRE; hand [email protected] and [email protected]; and PSNR, SSIM, and LPIPS. PCK counts missed hand detections as failures, whereas PCKm scores matched hands only; see Sec. C.1 for the evaluation protocol. Dashed lines indicate fitted trends.

Figure 7: Human-to-robot transfer scaling. PSNR, SSIM, and LPIPS on

\Psi_{0}

[ 117 ] , Xp-Teleop, and

\omega

-HOME [ 65 ] versus human-video pre-training hours (log scale); each point is a pre-training checkpoint post-trained on the target benchmark before evaluation. Colors and marker shapes distinguish benchmarks; lines connect measured checkpoint results.

Figure 8: Cross-evaluation scaling on the DreamDojo [ 34 ] benchmark. PSNR, SSIM, and LPIPS on the In-lab, EgoDex [ 46 ] , and DreamDojo-HV evaluation sets versus human-video pre-training hours (log scale), under an identical GR-1 post-training budget; the largest-scale point uses Ropedia W0 (5B). Colors and marker shapes distinguish evaluation sets; lines connect measured checkpoint results.

Figure 9: Policy evaluation and world-model-based planning. (a,b) Predicted vs. ground-truth task-progress scores in simulation and on the real robot (dashed: perfect agreement). (c,d) Task-progress scores for individual policies and world-model-based candidate selection; Random selects one candidate rollout uniformly at random, so its expected score equals the mean over candidates.

Demos/task	Initialization	LIBERO	LIBERO-Plus
Demos/task	Initialization	LIBERO	Camera	Robot	Language	Lighting	Background	Noise	Layout	Overall
full ( $\approx$ 43)	FastWAM	96.95	17.32	44.32	67.47	79.96	52.04	38.66	60.92	50.06
	Ropedia W0	97.00	20.14	51.68	73.52	88.05	61.43	48.09	66.30	56.82
	Improvement	+0.05 (0.1%)	+2.82 (16.3%)	+7.36 (16.6%)	+6.05 (9.0%)	+8.09 (10.1%)	+9.39 (18.0%)	+9.43 (24.4%)	+5.38 (8.8%)	+6.76 (13.5%)
10	FastWAM	87.30	18.32	23.10	43.59	55.27	29.46	28.36	42.36	33.58
	Ropedia W0	92.20	10.94	24.26	50.94	70.30	40.43	21.99	49.84	36.71

Table 6: World–action modeling on LIBERO [ 71 ] and LIBERO-Plus [ 32 ] . Success rates (%) with FastWAM [ 131 ] or Ropedia W0 init (LIBERO-Plus held out); best bold. Improvement is the absolute gain, relative gain over FastWAM in gray.

Figure 10: World–action modeling. (a) Real-robot executions of the policy initialized from Ropedia W0, shown as four chronological pairs of recorded keyframes with different objects, containers, and placements. (b) Aggregate real-robot task progress for Ropedia W0 initialization, FastWAM [ 131 ] , and the $\Psi_{0}$ [ 117 ] policy.

Figure 11: Live teleoperation with Ropedia W0. On-site recordings show the headset user and the monitor displaying model-generated feedback. Five chronological keyframes show interactions with human (top) and robot (bottom) embodiments.

Appendix figures & tables27 assets

Supplementary material from the paper’s appendix.

Appendix

Figure S1: Data curation funnel for Xp-Data. Robot and human egocentric recordings pass through source screening, motion annotation, and quality control; gray bands mark data removed at each step, and the surviving streams form the unified corpus of

\sim

3,000 robot hours and

\sim

10,000 human hours.

Figure S2: Motion annotations on the egocentric human data. Projected 21-point hand skeletons are overlaid on each frame, with the two hands drawn in different colors; the inset shows the camera trajectory of the surrounding window.

Figure S3: Example frames from the robot data in Xp-Data. The grid spans dexterous five-finger hands, three-finger hands, and parallel grippers on single-arm, dual-arm, and humanoid platforms, across household, retail, industrial, and tabletop scenes.

Figure S4: Motion annotations on the robot data. Image-aligned skeletons combine embodiment-native FK projections with manual keypoint annotation and correction (Sec. 3 ), covering dexterous hands and grippers; the inset shows the camera trajectory of the surrounding window.

Figure S5: Distillation of Ropedia W0 into a few-step causal student. Top: causal distillation initializes the few-step autoregressive generator. Bottom: asymmetric DMD updates the causal generator $G_{\theta}$ using a frozen bidirectional real-score teacher and an online-trained bidirectional fake-score model. The diagram illustrates the A14B MoE backbone; the 5B model follows an analogous procedure with a dense backbone. LongLive subsequently extends the few-step model for long-video streaming (Fig. S6 ).

Figure S6: LongLive-based streaming with the distilled student. Successive segments reuse the previously generated visual context while accepting updated controls (skeleton, camera trajectory, and text). Each block is produced by the few-step causal student, so the per-segment cost is bounded while retained context carries information across segments, enabling real-time long-horizon rollout.

Model	Rollout	FPS $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Teacher (50-step)	6.3 s	3.0	15.95	0.528	0.248
	19 s		14.34	0.454	0.321
	32 s		14.13	0.438	0.348
	61 s		13.64	0.414	0.385
Student (4-step)	6.3 s	21.3	16.12	0.538	0.247
	19 s		14.70	0.468	0.312

Appendix

Table S1: Distillation quality–speed comparison (GR1, 5B). The 4-step student matches or exceeds the 50-step teacher at every rollout length (bold) while running

7\times

faster.

Figure S7: Wide-FOV skeleton branch, example 1 (four moments in time order; columns labeled above). Wide-FOV skeleton conditioning improves the alignment of the visible arm near the image boundary. Here the outer border-band keypoint error drops from 8.9 to 6.9 px, and the wide branch is closer to the ground truth on 91% of frames.

Figure S8: Wide-FOV skeleton branch, example 2 (four moments in time order; columns labeled above). When the arm extends beyond the camera view, wide-FOV skeleton conditioning improves the prediction of its visible portion. Here the border-band keypoint error drops from 14.0 to 11.7 px, with the wide branch closer to the ground truth on 93% of frames.

Method	EgoSuite (100)			OpenAoE (100)
Method	ATE $\downarrow$	PSNR $\uparrow$	FVD 16 $\downarrow$	ATE $\downarrow$	PSNR $\uparrow$	FVD 16 $\downarrow$
Cosmos3-Nano (16B)	1.122	19.75	111.5	2.740	17.56	139.9
Cosmos3-Super (64B)	1.322	19.55	110.4	2.740	17.16	132.7
OSCAR (2B)	2.586	16.87	173.7	4.374	15.54	221.8
RynnWorld-Teleop (5B)	2.642	17.98	182.4	4.444	15.90	200.3
DreamDojo (2B)	1.892	17.94	173.8	3.412	15.57	241.1

Appendix

Table S2: Per-dataset results on EgoSuite [ 68 ] and OpenAoE [ 66 ] . Each set contains 100 clips from Table 2 . ATE is in cm and PSNR in dB. Best and second-best bold and underlined.

Model / initialization	No post-training			Few-shot			Full-data
Model / initialization	Seen	Unseen	$\Delta$	Seen	Unseen	$\Delta$	Seen	Unseen	$\Delta$
$\Psi_{0}$
DreamDojo (14B)	14.90	13.99	0.91	14.85	14.02	0.83	14.13	13.43	0.70
Cosmos3-Nano (16B)	13.83	13.65	0.18	15.04	14.37	0.67	17.10	16.00	1.10
Cosmos3-Super (64B)	13.82	13.73	0.09	15.27	14.60	0.67	16.23	15.22	1.01

Appendix

Table S3: Seen/unseen breakdown on

\Psi_{0}

[ 117 ] ,

\omega

-HOME [ 65 ] , and Xp-Teleop (PSNR

\uparrow

\Delta

is the seen

-

unseen gap. Best bold, second-best underlined.

Model / initialization	Overall			Per-segment PSNR $\uparrow$
Model / initialization	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	seg. 1	seg. 6	seg. 12
DreamDojo (14B)	14.17	0.641	0.579	22.32	13.36	13.02
Cosmos3-Nano (16B)	13.91	0.573	0.576	25.57	12.75	11.11
Cosmos3-Super (64B)	13.74	0.579	0.570	25.13	12.58	11.14

Ropedia W0 (5B), pretrain	13.64	0.651	0.516	23.92	12.57	10.35

Appendix

Table S4: Long-horizon rollout on

\omega

-HOME [ 65 ] (

\approx

32 s, zero-shot; 12 chained windows). Overall metrics are frame-averaged; per-segment PSNR at

\approx

1/14/31 s shows degradation over the rollout. Best bold , second-best underlined .

Model / initialization	Hand following			Fidelity			Video quality
Model / initialization	Err. (px) $\downarrow$	Miss $\downarrow$	Err.@120 $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	B-Cons.	Aes. $\uparrow$	Dyn.
DreamDojo (14B)	62.6	77.1%	106.9	10.00	0.301	0.620	0.891	0.449	0.55
Cosmos3-Nano (16B)	55.8	50.9%	88.5	11.10	0.338	0.579	0.892	0.500	0.55
Cosmos3-Super (64B)	52.5	61.5%	94.0	10.76	0.302	0.587	0.884	0.478	0.50

Ropedia W0 (5B), midtrain	19.8	45.3%	65.2	10.83	0.336	0.573	0.892	0.469	0.70

Appendix

Table S5: Long-horizon rollout on Xp-Teleop (

\approx

30 s; 11 autoregressively chained windows, each restarting from the model’s own last frame). Best bold , second-best underlined ; Dyn. is best when closest to the ground-truth value (0.70).

Model	Variant	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Cosmos3	Nano (16B)	19.02	0.739	0.197
Cosmos3	Super (64B)	18.77	0.736	0.202
Ours 5B	Full	19.83	0.760	0.167
	w/o Plücker	19.71	0.755	0.178
	w/o skeletons	18.03	0.736	0.214
Ours A14B	Full	21.98	0.802	0.128

Appendix

Table S6: Cross-embodiment set (400 evaluation windows over eight embodiments). Cosmos3 [ 1 ] is fine-tuned on the same robot mixture used for Ropedia W0’s mid-training; neither model is post-trained on the evaluation windows. Each ablation variant is trained without one conditioning channel. Best bold, second-best underlined.

Data	Base	In-lab			EgoDex			DreamDojo-HV
Data	Base	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
1%	Official base	16.04	0.495	0.446	13.42	0.394	0.548	12.43	0.348	0.548
1%	Ropedia W0	20.85	0.764	0.213	20.57	0.778	0.208	19.05	0.760	0.217
10%	Official base	18.89	0.630	0.287	18.64	0.687	0.267	16.95	0.623	0.307
10%	Ropedia W0	21.02	0.786	0.206	20.78	0.801	0.199	19.29	0.787	0.207
100%	Official base	19.41	0.650	0.275	18.87	0.695	0.265	17.48	0.645	0.293

Appendix

Table S7: Effect of pretraining. Our pretrained base vs. the same architecture from public video-generation weights (“official base”), post-trained on 1%/10%/100% of the GR-1 data; best bold.

Metric	BWM + SR	Ropedia W0 (A14B)
EWMScore	61.01	66.92
Image Quality	51.54	53.57
Aesthetic Quality	43.09	41.04
JEPA Similarity	90.93	96.00
Dynamic Degree	17.49	25.48
Flow Score	8.49	15.75

Appendix

Table S8: WorldArena [ 99 ] Track 1 comparison against BWM [ 17 ] (RoboTwin 2.0 [ 23 ] , 1000 episodes; identical official pipeline and inputs for both). EWMScore is the mean of the 15 metrics; best in bold.

Panel of Figure 33 — Figure S9: Qualitative comparison on WorldArena (left: drawer + box; right: cup onto coaster). Each panel: five chronological frames (rows); columns are GT $|$ Ropedia W0 (A14B) $|$ BWM (Table S8 ). Ropedia W0 follows the commanded arm trajectory and completes the task like the ground truth (opening the drawer and placing the box, precisely seating the cup), whereas BWM’s arm drifts and under-moves (it never opens the drawer).

Figure S10: Robot-dataset comparisons (1/3). (a) Fourier: putting tape in a bag; (b) AgiBot: grasping a desk fan. These are held-out clips: Ropedia W0 uses mid-training and Cosmos3 uses fine-tuning on the robot mixture. Colored skeletons are supplied input conditions overlaid on the generated frames.

Figure S11: Robot-dataset comparisons (2/3). (a) Baihu: placing a hot dog in a bowl; (b) Dexora: arranging fruit. Columns compare the ground truth, Ropedia W0, Cosmos3-Nano, and Cosmos3-Super at aligned frames. The held-out-clip setting and input skeleton overlays follow Fig. S10 .

Figure S12: Robot-dataset comparisons (3/3). (a) RH20T: turning a steering wheel; (b) DROID: closing a door. These examples extend the comparison to gripper motion and interactions with articulated objects. The held-out-clip setting and input skeleton overlays follow Fig. S10 .

Figure S13: Zero-shot comparisons on $\omega$ -HOME (1/2). (a) Reaching into the upper fridge; (b) pushing a chair. The compared models have not been trained on $\omega$ -HOME. Rows show aligned frames from the same clip; colored skeletons are input conditions overlaid on each generated video.

Figure S14: Zero-shot comparisons on $\omega$ -HOME (2/2). (a) Putting an apple in a fruit basket; (b) sorting fruit from office supplies. Ground-truth frames provide a reference for the generated hand and object configurations. The setting and input skeleton overlays follow Fig. S13 .

Figure S15: $\omega$ -HOME [ 65 ] manipulation task (fruit-bucket arrangement). Left: zero-shot; right: after post-training. Ropedia W0 follows the commanded hands and matches the ground truth in both settings, whereas the baselines lose the arm or drift.

Figure S16: Few-shot data efficiency on GR-1 (In-lab). Both models are post-trained on 1% of the data; Ropedia W0 initializes from our pretrained base and Wan2.2-I2V-A14B [ 108 ] from public video-generation weights. Our base yields markedly sharper, more faithful predictions.

Figure S17: Few-shot data efficiency on GR-1 (EgoDex [ 46 ] ). Same 1% budget; initializing from our pretrained base preserves object and hand detail that the public base loses.

Figure S18: Post-training examples on $\omega$ -HOME. (a) Wiping a table; (b) putting a towel in a washing machine. Columns show ground truth and Ropedia W0 under the three settings labeled in the project-page demonstrations: zero-shot, few-shot, and one epoch of target-data post-training. Colored skeletons are overlaid input conditions. Each row uses the same frame index across settings.

Figure S19: Additional 32-second generated rollouts. (a) Collecting fruit from a closet; (b) picking up bok choy. Each sequence starts with the input image at 0 s, followed by generated frames sampled every 6.4 s, read left to right and then top to bottom. The camera and hand-pose insets display supplied conditions.

Embodied World Models

Embodied world models predict how environments change under an agent’s actions to support planning, learning, simulation, and control.

See all 159 papers

$Ψ_0$ : An Open Foundation Model Towards Universal Humanoid Loco-Manipulation
Mar 12, 2026Songlin Wei, Hongyi Jing, Boqian Li +12Robot Foundation Models Robot Learning
USC Physical Superintelligence (PSI) Lab · NVIDIA · WorldEngine
$ω$ -0: A Latent Predictive World Action Model for Concurrent Humanoid Loco-Manipulation
Aug 6, 2026Zhe Li, Zhenzhe Zhang, Yangyang Wei +8Humanoid Loco-Manipulation Language-Conditioned Robot Manipulation
MARS Lab, NTU · BAAI · PKU +1
$λ_0$ : Scaling Open-World Egocentric Human Data for Humanoid Loco-Manipulation
Oct 13, 2026Lambda-humanoid TeamHumanoid Loco-Manipulation Whole-Body Robotic Manipulation
Alibaba Token Hub and Amap, Alibaba Group

Mar 12, 2026J/K move · Enter open · S save

$Ψ_0$ : An Open Foundation Model Towards Universal Humanoid Loco-Manipulation

Songlin Wei, Hongyi Jing, Boqian Li, Zhenyu Zhao+11

USC Physical Superintelligence (PSI) Lab · NVIDIA · WorldEngine

Robot Foundation Models Robot Learning Humanoid Loco-Manipulation Robotic Control Data-Efficient Robot Learning VLMs for Robotics

Open paper AskPDF arXiv

We introduce

Ψ_0

(Psi-Zero), an open foundation model to address challenging humanoid loco-manipulation tasks. While existing approaches often attempt to address this fundamental problem by co-training on large and diverse human and humanoid data, we argue that this strategy is suboptimal due to the fundamental kinematic and motion disparities between humans and humanoid robots. Therefore, data efficiency and model performance remain unsatisfactory despite the considerable data volume. To address this challenge, \ours;decouples the learning process to maximize the utility of heterogeneous data sources. Specifically, we propose a staged training paradigm with different learning objectives: First, we autoregressively pre-train a VLM backbone on large-scale egocentric human videos to acquire generalizable visual-action representations. Then, we post-train a flow-based action expert on high-quality humanoid robot data to learn precise robot joint control. Our research further identifies a critical yet often overlooked data recipe: in contrast to approaches that scale with noisy Internet clips or heterogeneous cross-embodiment robot datasets, we demonstrate that pre-training on high-quality egocentric human manipulation data followed by post-training on domain-specific real-world humanoid trajectories yields superior performance. Extensive real-world experiments demonstrate that \ours\ achieves the best performance using only about 800 hours of human video data and 30 hours of real-world robot data, outperforming baselines pre-trained on more than 10

\times

as much data by over 40% in overall success rate across multiple tasks. We will open-source the entire ecosystem to the community, including a data processing and training pipeline, a humanoid foundation model, and a real-time action inference engine.

Abstract

Embodied World Models

Explore similar work

Ψ0Ψ_0Ψ0​: An Open Foundation Model Towards Universal Humanoid Loco-Manipulation

ωωω-0: A Latent Predictive World Action Model for Concurrent Humanoid Loco-Manipulation

λ0λ_0λ0​: Scaling Open-World Egocentric Human Data for Humanoid Loco-Manipulation

$Ψ_0$ : An Open Foundation Model Towards Universal Humanoid Loco-Manipulation

$ω$ -0: A Latent Predictive World Action Model for Concurrent Humanoid Loco-Manipulation

$λ_0$ : Scaling Open-World Egocentric Human Data for Humanoid Loco-Manipulation