DynamicWAM: Dual-Path Motion Conditioning for World–Action Models in Dynamic Manipulation

Abstract

Dynamic manipulation requires robots to infer target motion and respond promptly, yet existing World–Action Models (WAMs) typically condition only on the current frame and execute large backbones synchronously, limiting motion awareness and responsive control in dynamic scenes. We propose DynamicWAM, a compact WAM for dynamic object manipulation with dual-path motion conditioning. DynamicWAM introduces history-flow conditioning, encoding temporally aligned optical-flow frames alongside the current observation through a frozen pretrained video VAE to preserve spatial motion structure, while injecting kinematic descriptors of displacement, duration, velocity, and acceleration into the action expert to provide motion magnitude and timing.

The two complementary paths are fused through joint world–action attention. A distilled compact backbone and real-time chunking (RTC)-based asynchronous execution further enable responsive control. On DOMINO, DynamicWAM achieves a 38.2% success rate and a 53.2 manipulation score, outperforming all evaluated baselines. Across 12 real-world tasks spanning linear, circular, and compound target motion, it achieves a 46.7% average success rate, exceeding the strongest baseline by 22.9 percentage points.

Method

DynamicWAM extends a standard World–Action Model with dual-path motion conditioning. Temporally aligned history-flow frames are encoded together with the current multi-view observation through a frozen pretrained video VAE, preserving where motion happens and how it is organized in image space. In parallel, kinematic descriptors of displacement, interval duration, velocity, and acceleration are projected into kinematic tokens and injected into the action expert, restoring the motion magnitude and timing that per-frame flow rendering discards. A compact video expert distilled from Wan2.2-TI2V-5B and the action expert interact through joint world–action attention, coupling future-video prediction with action generation.

The observation history is split into K = 4 intervals whose endpoints are Δ = 4 policy steps apart. For each interval, dense Farnebäck optical flow from the fixed head camera passes a forward–backward consistency check; the masked flow field is rendered with direction as hue and magnitude normalized by the frame's own 99th percentile. Statistics of the same flow field, together with the exact interval duration, form a 12-D kinematic descriptor that is standardized with dataset-level statistics and projected into a kinematic token. Intervals missing at episode start are marked invalid and receive learned validity embeddings.

For real-robot deployment, DynamicWAM adopts Real-Time Chunking: the next action chunk is predicted while the current one is being executed, hiding inference latency through execution–inference overlap. DOMINO evaluation keeps the benchmark's native synchronous protocol, so the simulated comparison isolates motion conditioning itself.

Simulation Results

On all 35 tasks of DOMINO Level 1 in the clean dynamic setting — 100 accepted episodes per task, unseen instructions, native synchronous protocol — DynamicWAM reaches 38.2% success and a 53.2 manipulation score, outperforming all eleven evaluated baselines on both metrics: +8.9 SR and +10.7 MS over the strongest baseline, InternVLA-A1.5, and +21.0 SR over PUMA. At 173.7 ms per policy query, it is also about 3.2× faster than InternVLA-A1.5, giving a favorable trade-off between inference efficiency and dynamic-manipulation performance.

The two paths contribute separately and compound. History-flow conditioning lifts the static WAM from 22.7% to 27.2% SR; kinematic tokens alone reach 30.5% SR and 47.2 MS; combining both yields 38.2% SR and 53.2 MS — 7.7 SR points above the stronger single path. The largest gains appear on tasks that require accurately timed contact: on move_pillbottle_pad success rises from 42% to 55% to 91% (static → + flow → full), and on beat_block_hammer from 36% to 49% to 84%. Kinematic tokens matter most on place_phone_stand, where history flow alone achieves 32% against 80% for the full model.

Real-World Experiments

We build a 12-task real-robot suite in three motion regimes, four tasks each: linear motion from a remote-controlled vehicle, an inclined rail, a conveyor belt, or a self-propelled toy; uniform circular motion on a rotating turntable; and compound motion including random bouncing, simultaneous translation and rotation, and curved trajectories. All policies run on Franka Research 3 arms with CTAG2F90 or Robotiq 2F-140 parallel-jaw grippers, are fine-tuned on the same 1,200 demonstrations, and are evaluated for 20 trials per task.

DynamicWAM achieves the highest success rate at every motion level — 70.00% linear, 51.25% circular, 18.75% compound — averaging 46.67%, 22.92 points above the strongest external baseline, π_0.5; on compound motion, no external baseline records a single successful trial. The variants keep their simulation ordering: history flow and kinematic tokens independently lift the static WAM from 25.00% to 36.25% and 40.00%, and combining both paths reaches 46.67%.

Asynchronous execution adds a complementary gain on top of motion conditioning: RTC improves the average success rate from 42.08% to 46.67%, and the synchronous variant already outperforms every external baseline. Under background, lighting, and object-color shifts on the Level-1 tasks, DynamicWAM achieves the highest success rate under every shift, improving out-of-distribution success to 36.3% versus 17.5% for π_0.5.

BibTeX

@article{dynamicwam2026,
  title   = {DynamicWAM: Dual-Path Motion Conditioning
             for World-Action Models in Dynamic Manipulation},
  author  = {Lou, Yunfan and Gao, Hewen and Zhu, Xiyu and
             Qiao, Zhuoran and Han, Xuan and Yang, Yifan and
             Ye, Yifan and Yao, Boxian and Pang, Zhibo},
  journal = {Under review},
  year    = {2026}
}