UniWAM Technical Report: Unified Mobile Manipulation via Mixed-Stream World-Action Modeling and Manipulation Anchor Pose Supervision

Wei Xue; Keliang Liu; Mingzhang Cui; Jinhua Xie; Jinjie Wei; Jianan Hou; Jingcheng Lu; Lintao Wang

cs.ROSep 30, 2026

UniWAM Technical Report: Unified Mobile Manipulation via Mixed-Stream World-Action Modeling and Manipulation Anchor Pose Supervision

Authors: Wei Xue, Keliang Liu, Mingzhang Cui, Jinhua Xie, Jinjie Wei, Jianan Hou, Jingcheng Lu, Lintao Wang, +9 more

Organizations: Physical Superintelligence Lab, Fysics AI · College of Intelligent Robotics and Advanced Manufacturing, Fudan University

Abstract

Mobile manipulation requires precise navigation to a manipulation-ready pose followed by reliable object interaction. These two stages differ in action spaces and visual requirements, which complicates unified policy learning. In addition, collecting diverse real-world navigation data with explicit manipulation-ready pose supervision remains costly and difficult to scale. We introduce UniWAM, a unified mixed-stream world-action model with separate action encoders and output heads for navigation and manipulation, sharing a common backbone. This design supports joint representation learning on independently sampled navigation and manipulation data. UniWAM supports independent inference for either stream and batch-parallel inference for both. We further introduce Manipulation Anchor Pose (MAP) supervision for where to stop and how to orient for manipulation. An automated pipeline constructs MAP-Data from large-scale 3D scenes, yielding over 1.5 million episodes and 7,500 hours. MAP-Data provides per-frame target-object bounding boxes and image-plane MAP coordinates as auxiliary navigation supervision. Together with projected end-effector trajectories for manipulation, these prediction targets provide stream-specific image-plane supervision for action learning from egocentric observations. With large-scale MAP-Data, UniWAM outperforms the strongest external baselines on our MAP-Bench by 30.1% in position error and 44.0% in heading error. Across 24 real-robot tasks, UniWAM achieves leading results in MAP navigation and mobile manipulation, with competitive manipulation performance. We have released code, data, and benchmark.

Figures & tables

Figure 1 : Challenges of mobile manipulation and our approach. Existing unified policies either share one action representation across navigation and manipulation or separate the two only inside the action module, and they learn where to stop only implicitly from costly complete trajectories. UniWAM models navigation and manipulation as separate streams over a shared backbone, and MAP-Data supervises the manipulation-ready pose at scale with automatically generated trajectories.

Model	Action generator	Separate views	One behavior per sample	Video world model	Explicit terminal pose
$\pi_{0.5}$ [ 3 ]	single action expert for all degrees of freedom	✗	✗	✗	✗
GR00T N1.7 [ 2 ]	single action head for all degrees of freedom	✗	✗	✗	✗
Xiaomi-Robotics-1 [ 32 ]	single action head for all degrees of freedom	✗	✗	✗	✗
ABot-M0.5 [ 9 ]	action stream with mobility and manipulation sub-towers	✗	✗	✓	✗
MobileWAM [ 12 ]	action expert with routed locomotion and manipulation experts	✗	✗	✓	✗
CrossFormer [ 11 ]	shared transformer with embodiment-specific action heads	✓	✓	✗	✗

Table 1 : Comparison with recent unified mobile-manipulation models and cross-embodiment policies. The first three are VLA policies, the next two are world-action models, and the last two are cross-embodiment policies, which separate different robots rather than the navigation and manipulation phases of one mobile manipulator. Separate views: navigation and manipulation receive different camera inputs. One behavior per sample: a training sample can supervise one behavior alone without padding or masking. Video world model: the model predicts future video. Explicit terminal pose: the terminal base pose for manipulation is supervised directly.

Method	Output	Label source	Supervises approach	Scale
Reachability inversion [ 26 ]	feasible base poses	arm kinematics	✗	–
MoMa-Kitchen [ 43 ]	floor affordance map	simulation	✗	127K episodes
FloAff-Kitchen [ 45 ]	floor affordance map	simulation	✗	24.8K observations
GBPP [ 8 ]	scores of candidate base poses	rule labels and simulation trials	✗	180K rule and 12K simulation labels
N2M [ 6 ]	distribution over base poses	policy rollouts	✗	12 to 15 rollouts per task
Mobi- $\pi$ [ 36 ]	base pose for a given policy	test-time optimization	✗	–

Table 2 : Comparison of MAP-Data with methods for manipulation-ready base placement. Supervises approach: the labels cover the entire approach trajectory rather than only a terminal pose or a score for a given scene. Scale is the amount of labeled data reported in each paper, and a dash denotes methods that do not learn from a dataset, such as kinematic search or test-time optimization.

Figure 2 : Overview of MAP-Data and MAP-Bench. From large-scale indoor 3D scenes, S1 curates valid target objects, S2 annotates each target with a MAP and feasible observation poses, and S3 synthesizes approach trajectories with RGB video, actions, and MAP and box labels. MAP-Bench evaluates position and heading errors from held-out initial poses.

Figure 3 : Overview of UniWAM. Navigation and manipulation samples are drawn independently, encoded by stream-specific action encoders, and packed as separate batch rows through the shared video and action DiTs. Stream-specific heads predict actions and image-plane auxiliary targets, while the video branch predicts future frames. Right: attention masks with clean action prefixes for training and inference, and the camera-frame end-effector pose shared across embodiments.

Figure 4 : Mobile manipulation platform. A shallow-angle head camera with a wide field of view provides the navigation-stream input, and a steep-angle head camera with a narrow field of view, together with the wrist cameras, provides the manipulation-stream input. In Mobile Manipulation, the navigation stream also receives the wrist cameras. The two inputs form the navigation and manipulation rows of a mixed-stream batch, which UniWAM processes in one forward pass.

	$\pi_{0.5}$	GR00T N1.7	UniWAM-Base	UniWAM-Scaled
Initialization	$\pi_{0.5}$ base checkpoint	GR00T N1.7-3B checkpoint	Wan2.2-TI2V-5B	Wan2.2-TI2V-5B
Robot pretraining	yes	yes	none	none
Frozen modules	none	none	VAE, text encoder	VAE, text encoder
Task demonstrations	24 tasks	24 tasks	24 tasks	24 tasks
Additional data	none	none	none	MAP-Data and 402 tasks
Image-plane auxiliary outputs	✗	✗	✓	✓

Table 3 : Real-robot training settings. Steps and GPUs are given per model. Each baseline trains one MAP Navigation model per mobile robot and one model for the remaining tasks of each embodiment, whereas each UniWAM variant is one model for all tasks.

	$E_{\mathrm{pos}}$ (cm) $\downarrow$			Overall	$E_{\mathrm{head}}$ ( ∘ ) $\downarrow$			Overall
Model	Short	Mid	Long	Overall	Short	Mid	Long	Overall
Vision-language-action models (VLA)
GR00T N1.7 [ 2 ]	9.48	10.8	22.0	14.1	5.66	5.42	7.46	6.18
$\pi_{0.5}$ [ 3 ]	7.98	9.85	50.8	22.9	3.26	3.51	5.55	4.11
Vision-language navigation models (VLN)
Uni-NaVid [ 42 ]	17.3	40.0	87.2	48.2	7.98	15.0	23.4	15.5

Table 4 : MAP-Bench results over the evaluation split. Bold and underline mark the best and second-best results.

Model	MAP-Data targets	$E_{\mathrm{pos}}$ Short	$E_{\mathrm{pos}}$ Mid	$E_{\mathrm{pos}}$ Long	$E_{\mathrm{pos}}$ Overall	$E_{\mathrm{head}}$ Overall
UniWAM-Base	0%	7.71	8.85	21.0	12.5	3.24
UniWAM	25%	7.63	8.52	18.1	11.4	2.81
UniWAM	50%	7.52	8.03	15.3	10.3	2.43
UniWAM-Scaled	100%	7.48	7.86	14.2	9.85	2.30
$\pi_{0.5}$	0%	7.98	9.85	50.8	22.9	4.11
$\pi_{0.5}$	100%	7.42	9.45	47.8	21.6	4.07

Table 5 : MAP-Bench errors when training with a fraction of the MAP-Data targets under a fixed compute budget. The second column is the fraction of MAP-Data targets from which the MAP-Data samples are drawn. Whenever MAP-Data is used, it is mixed with the MAP-Bench training split at a fixed 7:3 ratio, so all runs with MAP-Data see the same number of MAP-Data samples. 0% denotes training on the MAP-Bench split alone. UniWAM at 0% and 100% are UniWAM-Base and UniWAM-Scaled, and

\pi_{0.5}

uses the same data and budget. Position errors are in centimeters and heading errors in degrees.

	MAP Navigation		Mobile Manipulation			Manipulation
Model	$E_{\mathrm{pos}}$ (cm) $\downarrow$	$E_{\mathrm{head}}$ ( ∘ ) $\downarrow$	$E_{\mathrm{pos}}$ (cm) $\downarrow$	$E_{\mathrm{head}}$ ( ∘ ) $\downarrow$	SR (%) $\uparrow$	SR (%) $\uparrow$
$\pi_{0.5}$ [ 3 ]	23.3	5.15	26.2	5.72	62.5	71.4
GR00T N1.7 [ 2 ]	18.1	6.55	20.4	7.20	39.4	36.8
UniWAM-Base	15.1	4.91	17.1	5.30	63.1	44.3
UniWAM-Scaled	11.7	4.25	13.2	4.56	80.0	68.9

Table 6 : Real-robot results averaged over the tasks in each category. Bold and underline mark the best and second-best results.

MAP Navigation $E_{\mathrm{pos}}$ (cm) $\downarrow$ / $E_{\mathrm{head}}$ ( ∘ ) $\downarrow$
ID	Task	$\pi_{0.5}$	GR00T N1.7	UniWAM-Base	UniWAM-Scaled
1	Water dispenser	18.9 / 4.83	19.7 / 5.47	13.9 / 4.23	10.3 / 3.86
2	Humanoid robot	24.7 / 5.17	18.4 / 6.57	15.3 / 4.84	12.6 / 4.47
3	TV cabinet	20.2 / 4.56	14.8 / 5.74	12.1 / 3.76	8.94 / 3.58
4	Tool cabinet	25.3 / 5.63	20.2 / 7.18	16.5 / 7.27	11.7 / 4.06
5	Cushion	21.8 / 4.69	15.6 / 5.93	13.2 / 4.34	9.73 / 3.74

Table 7 : Per-task real-robot results with 40 trials per task on each embodiment. MAP Navigation cells report position error and heading error averaged over the Cobot Magic and the AgiBot G2. Mobile Manipulation cells additionally report the success rate in percent, and Manipulation cells report only the success rate. Avg. is the unweighted mean over the tasks in each category. Bold and underline mark the best and second-best values for each metric.

Reference: trained on real demonstrations
Model	Training data	$E_{\mathrm{pos}}$ (cm)	$E_{\mathrm{head}}$ ( ∘ )
$\pi_{0.5}$	MAP-Data only	55.1	11.4
GR00T N1.7	MAP-Data only	35.0	12.9
FastWAM	MAP-Data only	30.8	8.61
UniWAM-Sim	MAP-Data only	18.7	6.34
$\pi_{0.5}$	real task data	23.3	5.15

Table 8 : Zero-shot real-robot MAP Navigation of models trained only on MAP-Data, averaged over the Cobot Magic and the AgiBot G2. The bottom rows are models trained on real demonstrations, taken from Table 6 .

Model	Mobile Manipulation SR (%)	Manipulation SR (%)
$\pi_{0.5}$	62.5 [54.8, 69.6]	71.4 [65.9, 76.4]
GR00T N1.7	39.4 [32.1, 47.1]	36.8 [31.4, 42.6]
UniWAM-Base	63.1 [55.4, 70.2]	44.3 [38.6, 50.1]
UniWAM-Scaled	80.0 [73.1, 85.5]	68.9 [63.3, 74.1]
w/o navigation auxiliary supervision	56.9 [49.1, 64.3]	43.6 [37.9, 49.4]
w/o manipulation auxiliary supervision	58.8 [51.0, 66.1]	39.3 [33.7, 45.1]

Table 9 : Real-robot success rates with 95% Wilson intervals, pooled over 160 Mobile Manipulation trials and 280 Manipulation trials per method, matching the category means of Table 6 . Brackets give the interval bounds.

Figure 5 : Unified mobile-manipulation architectures compared in Table 10 : (a) unified action tokens, (b) separate navigation and manipulation action DiTs, (c) streams concatenated along the sequence axis, and (d) Mixed-Stream (ours), with streams in separate batch rows of shared video and action DiTs and independent stream sampling.

	Alternative architectures			(d) Mixed-stream
Metric	(a) Unified tokens	(b) Sep. experts	(c) Seq. concat.	Joint	Indep.
$E_{\mathrm{pos}}$ (cm) $\downarrow$	22.0	19.7	22.1	18.6	17.1
$E_{\mathrm{head}}$ ( ∘ ) $\downarrow$	7.49	6.06	6.78	5.87	5.30
SR (%) $\uparrow$	46.9	54.4	60.0	60.0	63.1

Table 10 : Architecture and sampling comparison on the four Mobile Manipulation tasks. (a) to (c) use joint sampling. (d) uses joint or independent sampling, and the latter is UniWAM-Base as reported in Table 6 .

	$E_{\mathrm{pos}}$ (cm) $\downarrow$			Overall	$E_{\mathrm{head}}$ ( ∘ ) $\downarrow$			Overall
Variant	Short	Mid	Long	Overall	Short	Mid	Long	Overall
UniWAM-Base	7.71	8.85	21.0	12.5	2.37	3.06	4.29	3.24
Velocity actions $(v,\omega)$	8.30	9.57	30.9	16.3	3.03	3.29	6.19	4.17
w/o causal mask	8.83	10.1	29.0	16.0	3.21	3.50	5.49	4.07
w/o target box	7.74	9.29	29.5	15.5	2.43	3.02	5.14	3.53
w/o image-plane MAP	8.05	8.40	22.5	13.0	2.63	2.82	4.72	3.39

Table 11 : Ablations of UniWAM-Base on MAP-Bench. The first row is the full model.

	MAP Navigation		Mobile Manipulation			Manipulation
Variant	$E_{\mathrm{pos}}$ (cm) $\downarrow$	$E_{\mathrm{head}}$ ( ∘ ) $\downarrow$	$E_{\mathrm{pos}}$ (cm) $\downarrow$	$E_{\mathrm{head}}$ ( ∘ ) $\downarrow$	SR (%) $\uparrow$	SR (%) $\uparrow$
UniWAM-Base	15.1	4.91	17.1	5.30	63.1	44.3
w/o navigation auxiliary supervision	17.6	6.84	19.9	7.21	56.9	43.6
w/o manipulation auxiliary supervision	15.4	5.18	17.8	5.61	58.8	39.3
w/o independent sampling	16.8	5.62	18.6	5.87	60.0	41.8

Table 12 : Ablations of UniWAM-Base on real robots. The first row is the full model.

	MAP Navigation		Mobile Manipulation			Manipulation
Tokens per bank	$E_{\mathrm{pos}}$ (cm) $\downarrow$	$E_{\mathrm{head}}$ ( ∘ ) $\downarrow$	$E_{\mathrm{pos}}$ (cm) $\downarrow$	$E_{\mathrm{head}}$ ( ∘ ) $\downarrow$	SR (%) $\uparrow$	SR (%) $\uparrow$
0	16.2	5.38	17.8	5.48	61.9	43.6
4, ours †	15.1	4.91	17.1	5.30	63.1	44.3
8	14.8	5.02	17.3	5.24	63.8	44.6

Table 13 : Real-robot results for the number of learnable tokens per bank.

\dagger

marks the default.

Target frame	Cobot Magic	Dual Franka	AgileX single arm	Mean
Robot base	55.0	28.1	62.5	48.5
Camera, ours	63.1	32.5	70.0	55.2

Table 14 : Manipulation target frame in UniWAM-Base under multi-embodiment co-training. Values are success rates in percent. Cobot Magic uses the Mobile Manipulation tasks, the dual Franka uses Manipulation tasks 1 to 3 and 7, and the AgileX single arm uses Manipulation task 6. Mean is the unweighted mean over the three reported embodiment and task groups.

Figure 6 : Real-robot executions of UniWAM. A single UniWAM model performs MAP Navigation, Mobile Manipulation, and Manipulation across four embodiments. The MAP Navigation examples dock at targets of different categories and sizes, the Mobile Manipulation examples show the transition from approach to manipulation, and the Manipulation examples show object sorting, test-tube insertion, and cup stacking.

Figure 7 : MAP-Bench trajectories of all evaluated methods on 24 randomly selected scenes. Subcaptions give the target category and initial distance. Stars mark the reference MAP, circles the start pose, and arrowheads the terminal heading. In the last two rows, every episode starts at least 3.5 m from the MAP, which falls in the long range.

Figure 8 : Navigation auxiliary predictions of UniWAM on MAP-Bench, overlaid over consecutive frames of each episode. Magenta boxes are the predicted target boxes, and cyan points are the predicted MAP positions projected into the image, both predicted jointly with the actions by the navigation stream of UniWAM at every step of the approach.

Figure 9 : Alignment between predicted 2D end-effector tracks and executed 3D actions on several tasks and two embodiments. Curves show the tracks of the two arms predicted in the previous action chunk, and circles mark the positions reached after that chunk.

Panel of Figure 10 — Figure 10 : Failure case in the color-sorting task. From left to right, a pink and a blue cube stand next to each other, and the gripper descends toward the point between them and closes on the empty gap instead of a cube.

Explore similar work

MobileWAM: Bridging World Action Models to Mobile Manipulation with Chain-of-Foresight
Aug 5, 2026Zehua Fan, Junjie He, Wenxuan Song +14World Models Kinematic Priors
Institute for AI Industry Research (AIR), Tsinghua University · 2Shanghai Jiao Tong University · 3The Hong Kong University of Science and Technology (Guangzhou) +8
ABot-M0.5: Unified Mobility-and-Manipulation World Action Model
Jul 1, 2026Ronghan Chen, Yandan Yang, Zuojin Tang +18Efficient World-Action Model Recent World-Action Models
AMAP CV Lab Alibaba Group
UMR: Universal Manipulation Representation
Sep 28, 2026Song Liu, Linyi Li, Yanshun Zhao +11Cross-Embodiment Transfer Unified Representation
University of Science and Technology of China, Hefei, China. · Suzhou Artificial Intelligence Laboratory, Suzhou, China.

Aug 5, 2026, cs.CVJ/K move · Enter open · S save

MobileWAM: Bridging World Action Models to Mobile Manipulation with Chain-of-Foresight

Zehua Fan, Junjie He, Wenxuan Song, Xi Wang+13

Institute for AI Industry Research (AIR), Tsinghua University · 2Shanghai Jiao Tong University · 3The Hong Kong University of Science and Technology (Guangzhou)+8

World Models Kinematic Priors Foresight Device

Open paper AskPDF arXiv

World action models (WAMs) built on video generation backbones are a rising recipe for robot learning, yet remain confined to tabletop manipulation. Mobile manipulation demands simultaneous locomotion and whole-body manipulation amid scene-scale dynamics, yet is still dominated by dynamics-blind visual encoders with hand-crafted coordination. We bridge this gap with MobileWAM, a mixture-of-transformers architecture that fuses a pretrained video diffusion transformer with a lightweight action expert through layerwise joint attention, translating internet-scale motion priors into whole-body control. To reconcile the heterogeneous dynamics of moving and manipulating, each feed-forward layer of the action expert becomes a three-expert mixture of shared, locomotion, and manipulation experts, softly routed by the motion intent in the action tokens. To densify supervision, we further propose Chain-of-Foresight (CoF): intermediate representations sequentially predict a chain of future latent chunks, each step conditioned on its predecessor. CoF pairs naturally with our decoupled video--action denoising scheme. At deployment, the WAM serves as a pure current-frame encoder; foresight acts only through gradients, so at inference the foresight chain and video generation are discarded, leaving only policy-level cost. MobileWAM surpasses state-of-the-art mobile manipulation policies on ManiSkill-HAB and fine-tunes to a real ARX Lift2 mobile manipulator across diverse tasks with strong generalization. Code will be released soon.

Abstract

Explore similar work

MobileWAM: Bridging World Action Models to Mobile Manipulation with Chain-of-Foresight

ABot-M0.5: Unified Mobility-and-Manipulation World Action Model

UMR: Universal Manipulation Representation