BridgeWAM

BridgeWAM in the real world

Four everyday tasks. Real-world manipulation.

Successful robot rollouts across different objects and settings.

LaptopSuccessful rollout

Abstract

Existing World Action Models (WAMs) commonly use Mixture-of-Transformers (MoT) to inject video tokens or their key–value representations directly into action attention. This dense coupling implicitly assumes natural alignment in a shared space and exposes the Action Expert to irrelevant visual information, limiting inference efficiency and generalization. We introduce Bridge-of-Experts (BoE), a compact interface that connects a Video Expert and an Action Expert through 32 Latent Bridge Queries (LBQs), and instantiate it in BridgeWAM. The LBQs aggregate observation, language, and robot-state information within the Video Expert and provide the Action Expert's conditioning. Joint video and action supervision trains this interface, with action gradients flowing through the LBQs into the Video Expert. BridgeWAM decouples expert depths, pairing a 30-layer Video Expert with a two-layer action head, and reuses the bridge throughout action denoising without generating future video at inference. BridgeWAM achieves 98.25% success on LIBERO and 72.24% in the LIBERO-Plus suite-level evaluation. Compared with Fast-WAM, it reduces FLOPs per inference by 28.72% in single-sample profiling and lowers inference latency to 77.1 ms in real-world robot experiments, a 2.47× speedup. These results suggest that BoE learns cross-modal alignment through joint video and action generation rather than assuming it in a shared attention space, allowing action supervision to guide the Video Expert toward representations better suited for task-relevant perception, action prediction, and closed-loop control across changing visual conditions.

Overview of BridgeWAM

BridgeWAM overview: LIBERO-Plus success and parameter count, the Latent Bridge Query architecture, and real-world results

Left: LIBERO-Plus success rates and model parameter counts. Center: Latent Bridge Queries connect a 30-layer Video Expert to a two-layer Action Expert, replacing the dense video K/V interface of MoT. Right: real-world manipulation tasks with success rates and measured inference latency.

Comparison of the BoE and MoT architectures

BoE uses Latent Bridge Queries (LBQs) to transfer action-relevant visual features between experts. Unlike MoT's matched-depth coupling, BoE decouples expert depths, enabling BridgeWAM to use only two Action DiT layers while retaining the full Video Expert for rich visual representation learning.

Model Architecture

BridgeWAM keeps the Video Expert and Action Expert modality-specific and couples them only through 32 Latent Bridge Queries (LBQs). Inside the 30-layer Video DiT, the LBQs are inserted from the first layer and, at every block, attend to the current-observation tokens (and to each other) to aggregate action-relevant visual and physical information. Crucially, the LBQs never attend to future-video tokens, so the same information boundary holds at training and inference — at test time only the current frame is processed.

After the final Video DiT layer, the LBQs are projected into the Action Expert space and read by a lightweight two-layer alternating cross–self-attention action head: a single cross-attention block retrieves information from the LBQs, followed by one self-attention block that reasons over the action chunk (“read once, reason once”). Video flow-matching and action flow-matching are optimized jointly, so action supervision propagates back through the LBQs and shapes the Video Expert toward representations better suited for control.

Results: LIBERO, LIBERO-Plus & LIBERO-Pro

Results on LIBERO and LIBERO-Pro. We report task success rates (%) on the Spatial, Object, Goal, and Long-Horizon suites, together with their average.

Method	Embodied PT.	LIBERO					LIBERO-Pro (zero-shot)
Method	Embodied PT.	Spatial↑	Object↑	Goal↑	Long↑	Avg.↑	Spatial↑	Object↑	Goal↑	Long↑	Avg.↑
OpenVLA	✓	84.7	88.4	79.2	53.7	76.5	40.5	32.0	25.0	20.5	29.5
π₀	✓	96.8	98.8	95.8	85.2	94.1	52.0	33.2	54.6	33.2	43.2*
π_0.5	✓	98.8	98.2	98.0	92.4	96.9	55.6	52.2	48.0	57.0	53.0*
LingBot-VA	✓	98.5	99.6	97.2	98.5	98.5	–	–	–	–	–
Motus	✓	96.8	99.8	96.6	97.6	97.7	–	–	–	–	–
World Action Model
Fast-WAM	✗	97.8	99.4	95.4	94.6	96.8	57.6	49.2	58.1	40.8	51.4
Faster-WAM	✗	99.6	99.8	98.2	98.2	99.0	59.6	52.8	59.6	45.8	54.5
BridgeWAM (Ours)	✗	99.0	100.0	97.4	96.6	98.3	57.8	55.1	58.4	43.5	53.7

* Reproduced using the same hyperparameter settings as the official LIBERO-Pro repository.

Zero-shot results on LIBERO-Plus. Task success rates (%) under seven types of perturbations and their average. “Embodied PT.” indicates whether embodied pre-training is used.

Method	Embodied PT.	LIBERO-Plus (zero-shot)
Method	Embodied PT.	Camera↑	Robot↑	Lang.↑	Light↑	Backg.↑	Noise↑	Layout↑	Avg.↑
Vision Language Model
OpenVLA	✓	0.6	3.9	38.7	14.2	38.4	5.4	42.5	19.7
OpenVLA-OFT	✓	56.4	31.9	79.5	88.7	93.3	75.8	74.2	69.6
π₀	✓	13.8	6.0	58.8	85.0	81.4	79.0	68.9	53.6
World Action Model
Fast-WAM	✗	15.4	44.4	68.5	79.0	52.6	37.3	61.5	49.5
OpenWAM-α	✗	33.8	76.1	88.0	97.0	87.1	39.8	77.5	69.2
Faster-WAM	✗	53.8	71.6	94.7	96.3	61.3	63.6	79.1	73.6
BridgeWAM (Ours)	✗	56.5	60.5	85.1	92.2	69.2	73.6	75.7	72.2

Results: RoboTwin 2.0

Fine-grained task comparison on RoboTwin 2.0 Simulation (Clean vs. Randomized).

Task	π₀		π_0.5		Motus		Motus (WAN2.2)		LingBot-VA		Fast-WAM		Faster-WAM		BridgeWAM (Ours)
Task	C	R	C	R	C	R	C	R	C	R	C	R	C	R	C	R
Adjust Bottle	99	95	100	99	89	93	99	97	90	94	100	100	100	99	100	99
Blocks Ranking RGB	80	63	92	85	99	97	92	88	99	98	100	100	100	100	99	99
Click Bell	71	48	99	66	100	100	100	100	100	100	100	100	100	100	99	99
Grab Roller	98	94	100	100	100	100	100	100	100	100	100	100	100	100	100	100
Handover Mic	97	97	98	97	78	63	98	95	94	96	99	100	100	99	100	100
Lift Pot	80	72	96	85	96	99	90	87	100	99	100	100	100	100	100	100
Shake Bottle Horizontally	98	92	99	99	100	98	100	97	100	99	100	100	100	100	100	100
Shake Bottle	94	91	99	97	100	97	100	96	100	97	100	100	100	100	100	100
Move Pillbottle Pad	67	46	84	61	93	96	83	83	99	99	100	99	98	99	99	100
Move Playingcard Away	74	65	96	84	100	96	50	47	100	99	100	100	100	100	98	100
… (50 tasks) …
Click Alarmclock	77	68	98	89	100	100	100	99	99	100	100	100	99	100	97	99
Dump Bin Bigbin	88	83	92	97	95	91	94	96	89	96	97	96	97	98	98	97
Open Laptop	71	81	90	96	95	91	89	89	92	94	98	100	95	99	99	100
Place Container Plate	97	92	99	95	98	99	97	100	99	97	96	100	98	97	99	100
Place Burger Fries	81	76	94	87	98	98	95	90	97	95	96	99	98	97	98	98
Place Object Stand	82	68	91	85	98	97	91	93	99	96	90	94	95	96	96	98
Place Phone Stand	49	53	81	81	87	86	80	78	97	97	97	99	100	96	98	95
Place Shoe	76	76	92	93	99	97	95	92	98	98	96	99	97	98	97	96
Stack Bowls Two	94	95	95	96	98	98	98	98	94	98	92	98	98	96	99	97
Stack Blocks Two	93	79	97	100	100	98	96	94	100	98	100	100	100	100	100	99
Stack Bowls Three	77	75	77	71	79	87	90	74	86	83	80	81	83	78	88	90
Average (%)	65.92	58.40	82.74	76.76	88.66	87.02	77.56	77.00	92.90	91.50	91.88	91.78	92.78	92.26	89.84	90.22

C = Clean scenes, R = Randomized scenes; 100 trials per task and condition. Averages over all 50 RoboTwin 2.0 tasks.

Real-World Bimanual Manipulation

We deploy BridgeWAM on the PiPER X (SongLing) real-robot platform and evaluate four manipulation tasks — closing a laptop and picking-and-placing a drink, a pen, and a mouse — together with their out-of-distribution (OOD) settings that vary object height, type, color, position, or count. Using task-specific staged scoring, BridgeWAM reaches an average success rate of 68.92%, compared with 59.83% for Fast-WAM, while generating each action chunk in 77.1 ms on the robot. This shows that the LBQ-mediated video–action interface transfers beyond simulation and stays stable under real-world visual variation, observation noise, and execution disturbances.

Real-world manipulation tasks, success rates, and inference latency on the PiPER X platform — Real-world manipulation tasks, success rates, and inference latency. View full-resolution PDF ↗

Real-world rollout comparison

FastWAM on the left, BridgeWAM on the right. Play each task pair together or use the individual video controls.

FastWAM

BridgeWAM Ours

FastWAM

BridgeWAM Ours

FastWAM

BridgeWAM Ours

FastWAM

BridgeWAM Ours

Video presentation notes

Clips play at their original speed and are paired by task and filename index, not by frame-aligned robot trajectories. Clip lengths and fields of view can differ. FastWAM drink 0 omits its first 5 seconds; FastWAM mouse 0 and BridgeWAM drink 0/1/2 are cropped by 15% from the left. No additional crop, speed change, or time trim is applied for this comparison.

Drink has three paired settings. Pen and mouse show setting 0, the available matching FastWAM clip; additional BridgeWAM settings remain in the top-of-page showcase.

Ablation Studies

Action-head depth ablation on LIBERO and LIBERO-Plus. Report task success rates (%).

Method	LIBERO					LIBERO-Plus (zero-shot)
Method	Spatial↑	Object↑	Goal↑	Long↑	Avg.↑	Spatial↑	Object↑	Goal↑	Long↑	Avg.↑
FastWAM	97.8	99.4	95.4	94.6	96.80	52.21	69.06	37.32	40.06	49.54
BridgeWAM (Ours)	99.0	100.0	97.4	96.6	98.25	77.64	84.47	60.29	67.17	72.24
BridgeWAM 4 Layers	97.2	100.0	97.2	96.6	97.75	67.44	84.63	54.30	60.26	66.56
BridgeWAM 8 Layers	97.4	99.8	97.6	96.8	97.90	70.69	83.76	49.83	58.44	65.50
BridgeWAM 16 Layers	98.8	99.4	96.2	93.4	96.95	67.94	80.70	52.37	62.13	65.66
BridgeWAM 30 Layers self-cross head	97.8	100.0	97.6	94.8	97.55	69.40	78.04	60.09	64.31	67.89

Interface and video–action coupling ablations on LIBERO and LIBERO-Plus. We report task success rates (%).

Method	LIBERO					LIBERO-Plus (zero-shot)
Method	Spatial↑	Object↑	Goal↑	Long↑	Avg.↑	Spatial↑	Object↑	Goal↑	Long↑	Avg.↑
FastWAM	97.8	99.4	95.4	94.6	96.80	52.21	69.06	37.32	40.06	49.54
BridgeWAM (Ours)	99.0	100.0	97.4	96.6	98.25	77.64	84.47	60.29	67.17	72.24
BridgeWAM Frozen Video DiT	87.8	97.0	84.8	72.4	85.50	32.22	38.52	15.44	16.36	25.48
BridgeWAM Use MoT (text_state)	98.0	99.4	97.4	96.4	97.80	71.57	80.70	36.09	54.78	60.48
BridgeWAM text_state_LBQ input	96.4	99.2	97.0	96.2	97.20	64.70	79.98	44.11	49.07	59.29
BridgeWAM Joint	97.4	99.6	96.6	96.8	97.60	76.52	72.12	46.51	59.59	63.41
BridgeWAM IDM	98.8	97.8	99.0	96.6	98.05	76.23	71.49	58.55	64.19	67.45

Inference Efficiency

Inference efficiency on LIBERO. Both methods use a 30-layer Video DiT and are profiled on the same H20Z GPU in bfloat16 with an action horizon of 32 and 10 denoising steps.

Method	Architecture		Runtime-Shape FLOPs (T)			Acceleration		Peak Mem. (MiB)↓
Method	Video	Action	Video↓	Action↓	Total↓	Latency (ms)↓	Speedup↑	Peak Mem. (MiB)↓
Fast-WAM	30	30	1.005	1.043	2.048	405.91	1.00×	23,773.86
Faster-WAM	30	30	2.745	0.476	3.221	321.16	1.26×	22,962.00
BridgeWAM (Ours)	30	2	1.432	0.028	1.460	88.05	4.61×	21,912.21

Action-Conditioned Visual Routing

To visualize which visual information is preferentially accessed by the Action Expert, we construct an action-conditioned visual-routing map. For BridgeWAM, we use the Action Expert's attention scores over the latent bridge queries as query-importance weights and apply them to the corresponding LBQ-to-video-patch attention maps from the Video Expert. For Fast-WAM, we directly aggregate the action-to-visual-token attention across all Video-DiT layers. The resulting patch-space maps reveal the visual information preferentially routed toward action prediction; they should be interpreted as routing visualizations rather than strict causal attributions.

Action-conditioned visual routing maps. The top row shows Fast-WAM attention averaged across Video-DiT layers, while the bottom row shows the LBQ-mediated routing at the final Video-DiT layer of BridgeWAM. The columns cover variations in background texture, lighting, camera viewpoint, robot initial state, object layout, and sensor noise.

Conclusion

We presented Bridge-of-Experts (BoE), a differentiable Latent Bridge Query interface that aligns video representations with action prediction and enables selective retrieval by the Action Expert. Reusing the compact bridge supports independent expert depths and efficient conditioning, while our experiments validate its promise for robust, longer-horizon control across diverse settings.

BridgeWAM: Does Shared Attention Align Physical World Representations with Actions?

Four everyday tasks. Real-world manipulation.

Abstract

Overview of BridgeWAM

Comparison of the BoE and MoT architectures

Model Architecture

Results: LIBERO, LIBERO-Plus & LIBERO-Pro

Results: RoboTwin 2.0

Real-World Bimanual Manipulation

Real-world rollout comparison

Ablation Studies

Inference Efficiency

Action-Conditioned Visual Routing

Conclusion

Acknowledgements

BibTeX

BridgeWAM: Does Shared Attention Align Physical World Representations with Actions?

Four everyday tasks. Real-world manipulation.

Abstract

Overview of BridgeWAM

Comparison of the BoE and MoT architectures

Model Architecture

Results: LIBERO, LIBERO-Plus & LIBERO-Pro

Results: RoboTwin 2.0

Real-World Bimanual Manipulation

Real-world rollout comparison

01Close the laptop

02Pick and place drink

03Pick and place pen

04Pick and place mouse

Ablation Studies

Inference Efficiency

Action-Conditioned Visual Routing

Conclusion

Acknowledgements

BibTeX