PaperlayerView exact revision on arXiv

Source-linked AI summary

AquaWAM: A Dynamics-aware World Action Model for Underwater Embodied Agents

Cunhao Zhu, Yifeng Wang, Dongliang Xu, Yunzhong Hou, Yue Yao, Chi Harold Liu

arXiv:2609.33299v2cs.ROcs.AI

TL;DR

Underwater robots must account for passive motion that persists after commands, which existing video-predicting WAMs are not designed to model explicitly. AquaWAM predicts compact sensor-based physical states, including passive dynamics, and reaches 72.6% success across 20 USIM tasks versus 51.7% for U0.

  • Problem

    Underwater motion persists beyond commands, while existing WAMs primarily predict visual observations and are not explicitly designed to capture passive dynamics.

    Source: §1 · p. 2

  • Method

    AquaWAM predicts compact sensor-measured physical states and learns commanded and passive dynamics, with cameras supplying goals and target poses.

    Source: §1 · p. 2, §3.1 · p. 4

  • Results

    72.6% success across 20 USIM tasks versus 51.7% for U0, with both fine-tuned on the same demonstrations under one protocol.

    Source: §1 · p. 2

  • Takeaways & Limitations

    AquaWAM’s compact model runs onboard, and the same predictor can provide velocity when the DVL fails.

    Source: §5 · p. 9

    All results are in simulation; real deployment adds effects the simulator only approximates, and would require safety validation beyond this paper’s scope.

    Source: §5 · p. 9, ETHICS STATEMENT · p. 10

Abstract

from arXiv · show

World Action Models (WAMs) are becoming increasingly important and useful for embodied intelligence, as they enable robots to anticipate the consequences of candidate actions before interacting with the physical environment. However, underwater robots are usually subject to passive dynamics, such as inertia, buoyancy, hydrodynamic drag, and persistent drift, which can continue to affect the vehicle even after an action is completed. Existing WAMs, which primarily predict action-conditioned visual observations, are not explicitly designed to capture such passive motion dynamics. In this paper, we present AquaWAM, the first World Action Model designed for underwater embodied agents. Instead of predicting future images, AquaWAM models both action-conditioned and passive physical dynamics, including the thruster dead band, the inertial glide that outlasts each command, and ambient currents. Specifically, it senses through the DVL, IMU, pressure sensor and joint encoders, while cameras supply only semantics for understanding goals and target pose. By modeling compact navigation states rather than high-dimensional visual observations, AquaWAM substantially reduces the model size and computational cost compared with conventional WAMs. Experimentally, AquaWAM achieves a 72.6% task success rate across 20 underwater tasks on the USIM benchmark, outperforming existing methods while making action decisions 2.7x faster than U0 on an NVIDIA Jetson AGX Orin. Our model also remains effective when some onboard sensor measurements are unavailable. For example, without DVL velocity measurements, our method still achieves a 61.6% success rate, compared with 39.4% for U0.

1 INTRODUCTION

Underwater robots must anticipate passive motion that persists beyond commands, while video-based WAMs are costly and indirectly represent motion. AquaWAM predicts compact physical states and achieves higher USIM success than U0 while making decisions faster.

1 INTRODUCTION

  • Underwater inertia, buoyancy, drag, and currents sustain motion beyond commands, while grasping requires centimeter-level positioning and small corrections can fall within the thruster dead band.DVL bottom lock may also fail, removing velocity measurements.Source: §1 · p. 2
  • Video-based WAMs are computationally costly on embedded platforms and indirectly represent future vehicle motion, especially under degraded visibility.The paper motivates direct physical-state prediction that remains lightweight for evaluating candidate actions onboard.Source: §1 · p. 2
  • AquaWAM predicts a 35-dimensional sensor-measured physical state, uses cameras only for goals and object poses, and learns passive effects from task-agnostic play as well as demonstrations.It uses 2.4M parameters for cruising and 2.6M near an object, versus 2.7B for U0; decisions are 2.7× faster on the same Jetson AGX Orin hardware.Source: §1 · pp. 2–3
  • 72.6% success across 20 USIM tasks versus 51.7% for U0, with both models fine-tuned on the same demonstrations under one protocol.The comparison is reported against U0, identified as the benchmark’s strongest model.Source: §1 · p. 2

2 RELATED WORK

AquaWAM differs from underwater control and VLA approaches by selecting actions from predicted physical outcomes and using one predictor for both action prediction and velocity estimation. It treats video as optional rather than a required world representation.

2.1 UNDERWATER STATE ESTIMATION AND CONTROL

  • Unlike approaches that separate control and state estimation, AquaWAM predicts candidate-action outcomes and uses the same predictor to provide velocity when the DVL fails.Earlier approaches include hydrodynamic control, thruster models, and learned velocity estimation.Source: §2.1 · p. 3

2.2 VISION-LANGUAGE-ACTION MODELS

  • VLA models primarily reproduce expert actions, and four of the six compared models issue 16-step chunks without intermediate outcome evaluation; U0 outperforms several compared VLAs on USIM.The cited comparison names OpenVLA, π0.5, and GR00T N1.5 as weaker than U0 when trained on the same demonstrations.Source: §2.2 · p. 3

2.3 WORLD MODELS

  • AquaWAM follows model-predictive action selection but replaces video with onboard-sensor physical state, exposing control-relevant quantities in a compact representation.Earlier video world-model work includes VPP and UVA; Figure 2 shows camera-supplied goals and object pose alongside candidate prediction.Source: §2.3 · p. 3, Figure 2 · p. 4

3 AQUAWAM: A DYNAMICS-AWARE WORLD ACTION MODEL

AquaWAM predicts candidate futures from compact sensor states, selects commands by their predicted outcomes, and reuses its predictor for navigation-state estimation when DVL velocity is unavailable. Cameras provide objectives and target pose rather than dynamics inputs.

3 AQUAWAM: A DYNAMICS-AWARE WORLD ACTION MODEL

  • A single predictor forecasts physical states under candidate commands, selects commands from those predictions, and replaces the DVL when it fails.It imagines candidates two seconds ahead and executes the first 0.5 seconds of the selected commands.Source: §3 · p. 4

3.1 PHYSICAL STATE AND PASSIVE DYNAMICS

  • The predictor uses a 19-dimensional vehicle state while cruising and a 35-dimensional state near an object, adding arm and gripper-relative object pose.The state includes DVL velocity and altitude, IMU angular rate and acceleration, pressure-sensor depth, and the last thruster command.Source: §3.1 · p. 4
  • A recurrent history encoder represents ambient current and thruster efficiency, while the predictor learns glide, dead band, buoyancy, and drag from demonstrations and task-agnostic play.After thrust stops, speed decays with a 2.1 s time constant; commands below about 0.12 change speed by less than 0.4 cm/s per second, while commands between 0.3 and 0.6 change it by about 4 cm/s per second.Source: §3.1 · p. 4

3.2 PREDICTION BEFORE ACTION

  • AquaWAM scores 128 candidate command sequences rolled out two seconds ahead, revises its choice every half second, and executes only the first half second of the winner.The second candidate round is drawn around the sixteen best first-round candidates; USIM supplies sub-goals and goal velocity.Source: §3.2 · p. 5

3.3 THRUST PREDICTION FOR FINE POSITIONING

  • For fine positioning, AquaWAM evaluates thrust pulses followed by coasting at their predicted endpoint, selecting a glide that ends within the grasp window at rest.Pulse magnitudes cross the dead band by construction; the hull provides centimeter-level positioning while the arm holds a ready pose.Source: §3.3 · p. 5

3.4 SEMANTIC DECODING FOR OBJECTIVES

  • Perception heads convert images and instructions into waypoint goals and target poses; camera images are not direct inputs to the dynamics predictor.The wrist-camera head also locates the destination box for transport.Source: §3.4 · p. 5

3.5 AQUAWAM AS A NAVIGATION-STATE ESTIMATOR

  • When DVL velocity is masked, three velocity heads estimate it from history, IMU readings, and commands; the estimate replaces DVL velocity in the dead-reckoned pose.A cumulative trust statistic bounds how far decisions may depart from the last DVL-based decision; trust becomes immediate after a known goal change.Source: §3.5 · p. 6

4 EXPERIMENTS

Across USIM’s 20 tasks, AquaWAM outperforms U0 under full sensing and retains higher success under DVL loss. The experiments examine task-level grasping, extreme conditions, component ablations, and inference cost.

4 EXPERIMENTS

  • The evaluation uses USIM’s 20 tasks and prescribed 700 trials per model, with shared cameras, commands, and judges across models.In the dropout condition, the DVL loses bottom lock partway through each trial and never recovers.Source: §4 · p. 6

4.1 MAIN RESULTS

  • AquaWAM succeeds in 72.6% of 700 full-sensing trials versus U0’s 51.7%, with gains across task families and higher grasp, transport, and tracking success.Navigation rises from 87.5% to 94.4%; grasping from 40.6% to 65.4%; transporting from 32.5% to 57.5%; and tracking from 70% to 100%.Source: §4.1 · pp. 6–7, Table 1 · p. 7

4.2 GRASP EXPERIMENTS

  • Across grasp objects, AquaWAM beats U0 in every task, and predicted thrust pulses succeed in 85.2% of 480 trials versus 17.9% for a hand-tuned positioning law.The law brings the gripper inside the closure window in only 24% of trials; cadence, closure timing, and pulse-grid alternatives do not explain the margin.Source: §4.2 · p. 7, Figure 5 · p. 7

4.3 WHEN THE DVL FAILS

  • With DVL loss, AquaWAM retains 61.6% success versus U0’s 39.4%, including 93.8% navigation and 80% tracking success on dead reckoning.As a takeover layer, its velocity estimate raises U0’s success from 39.4% to 49.7%; transporting after DVL loss remains low for both methods.Source: §4.3 · p. 8, Figure 6 · p. 8

4.4 EXPERIMENTS ON EXTREME CASES

  • Under perturbed worlds, AquaWAM outperforms U0 on every reported protocol, while stricter waypoint results vary by tolerance and route.With DVL loss, it passes 0.5 m waypoints in 45 of 80 gotos versus 28 for U0, and completes 39 of 40 inspections versus 7.Source: §4.4 · p. 8, Table 2 · p. 8

4.5 ABLATIONS

  • Ablations show that predicted rollouts, play data, a disturbance token, and AquaWAM velocity estimation each support performance where their modeled dynamics matter.Without imagination, grasping falls from 65.4% to 16.5% and transporting from 57.5% to 7.5%; without play data, navigation drops to 50.6%.Source: §4.5 · p. 8, Table 3 · p. 9

4.6 COMPUTATION COSTS

  • On Jetson AGX Orin, AquaWAM’s end-to-end decisions take 79.5 ms on navigation and 95.7 ms on grasping, 2.7× faster than U0’s 214 ms.The predictor has 2.4M–2.6M parameters; perception backbones add 86M each and their time is included in the decision figures.Source: §4.6 · p. 9

5 CONCLUSION

AquaWAM predicts a 35-dimensional sensor-based physical state and learns passive dynamics from task-agnostic play. The conclusion highlights onboard inference, grasping performance, and DVL-failure estimation, while noting that evaluation remains in simulation.

5 CONCLUSION

  • AquaWAM runs on a Jetson AGX Orin at 79.5 ms per decision, and its predicted thrust pulses grasp within a centimeter where a hand-tuned law fails.The same predictor can also replace a failed DVL, for AquaWAM itself and as a takeover layer for a VLA.Source: §5 · p. 9
  • All results are in simulation; real deployment adds effects the model only approximates, including sensor noise, latency, and intermittent DVL dropouts.Transport after DVL loss remains at 7.5%; a thruster fault absent from play leaves only 4 of 40 gotos successful, and wrist-camera perception yields 65.4% grasping versus 85.2% with benchmark object pose.Source: §5 · p. 9
  • A physical ROV deployment is future work; the authors propose adapting with unlabeled play data from the target hull and using visual cues to bound dead-reckoning drift.The paper states that deploying a learned controller on a real vehicle would require safety validation beyond its scope.Source: §5 · p. 9, ETHICS STATEMENT · p. 10

A THE PREDICTOR

AquaWAM predicts compact physical states from vehicle sensors and selects actions by evaluating predicted outcomes. Its appendix describes the state, prediction pipeline, scores, and hand-tuned comparison law.

A THE PREDICTOR

  • AquaWAM uses a 19-dimensional vehicle state while cruising and adds 16 dimensions near an object; its predictor forecasts every channel.The history encoder maps the last 16 steps to a 96-dimensional disturbance token; chained calls predict five steps each, with four calls covering two seconds.Source: A THE PREDICTOR · p. 14, Table 4 · p. 14
  • Every half second, AquaWAM generates candidate actions, predicts their outcomes, selects the lowest-scoring option, and executes its first half second.When DVL bottom lock is lost, it estimates velocity from history and dead-reckons pose; near an object it uses the full state and thrust pulses.Source: A THE PREDICTOR · p. 14

B SCORES AND THE HAND-TUNED LAW

  • The score weights are λg = 5, λu = 0.02, λs = 0.4, and λω = 0.5; task-specific speed caps are soft penalties.The goal term can outweigh the cap during tracking, allowing the vehicle to match the boat’s 0.44 m/s.Source: B SCORES AND THE HAND-TUNED LAW · p. 14
  • The pulse candidates vary thrust magnitude and duration, include a coast step, and use velocity-dependent directions and a hold option.The pulse score evaluates predicted endpoint error and velocity over the final six steps of the two-second horizon.Source: B SCORES AND THE HAND-TUNED LAW · pp. 14–15
  • The hand-tuned law shares AquaWAM’s task program and closure window, differing only in hull positioning; it approaches at 0.03 m/s and switches to grasp at 8 cm.Its closure conditions also require specified horizontal offsets, distance, yaw error, and a stationary hull.Source: B SCORES AND THE HAND-TUNED LAW · p. 15

C TASKS

The section identifies a table of 20 USIM tasks and trials per task.

C TASKS

  • Table 5 lists the twenty USIM tasks and trials per task.Source: Table 5 · p. 15

D PLAY DATA

The play data are unlabeled recordings of the vehicle rather than task demonstrations. Their differing commands expose dynamics that demonstrations alone do not reveal.

D PLAY DATA

  • Play recordings contain no task or success labels and use commands unlike expert demonstrations, exposing the thruster dead band and glide beyond commands.The play comprises three recordings from the same vehicle: exploration, a pass through the scenes, and a deployment log.Source: D PLAY DATA · p. 15

E BASELINES AND DVL DROPOUT

AquaWAM’s performance was evaluated under shared fine-tuning demonstrations and DVL dropout, including use of its velocity estimate as a takeover layer for U0.

E BASELINES AND DVL DROPOUT

  • Every Table 1 model was fine-tuned on the same demonstrations, with model-specific training and action-chunk protocols.U0 and GR00T N1.5 commit to 16-step chunks; OpenVLA and π0.5 act at every step.Source: E BASELINES AND DVL DROPOUT · pp. 15–16
  • Under DVL dropout, tracking fell from 20 to 16 of 20 trials, grasping lost at most 12 of 40 successes per object, and transport fell from 23 to 3 of 40.Goto, scan, and inspection were almost unaffected.Source: E BASELINES AND DVL DROPOUT · p. 16, Table 6 · p. 16
  • With AquaWAM’s velocity estimate replacing lost DVL measurements, U0 navigation rose from 73 to 155 of 160 trials and tracking from 1 to 12 of 20.Grasping fell from 202 to 181 of 480, while transport remained at 0 of 40.Source: E BASELINES AND DVL DROPOUT · p. 16, Table 7 · p. 16

F USIM-HARD

USIM-HARD evaluates models under seeded changes to task conditions and stricter scoring of recorded trajectories.

F USIM-HARD

  • Perturbations alter heading, goto endpoint tolerance, horizontal thruster output, or water turbidity, with 20 trials per cell.A thruster may be weakened to 50% or disabled; turbid water raises the Jerlov parameter from 0.15 to 0.50.Source: F USIM-HARD · p. 16
  • Each perturbation is applied before scene loading and seeded by task and trial, so models face the same change in a given trial.Source: F USIM-HARD · p. 16
  • Stricter judges rescore recorded trajectories without perturbing the scene, requiring every goto waypoint and specified inspection-node coverage.For inspections, every node must be within a 1 m ball and 90% of nodes must be covered.Source: F USIM-HARD · p. 16

G FAILURE CASES

The reported failure cases expose limits under DVL loss, a completely failed thruster, and degraded camera visibility.

G FAILURE CASES

  • The task-level breakdowns cover perturbed worlds and stricter judges, with 20 trials per cell for perturbed worlds and 40 goto or 20 inspection trials for stricter judges.Source: Table 8 · p. 17, Table 9 · p. 17
  • With DVL lost during transport, AquaWAM succeeds in 3 of 40 trials versus U0’s 0 of 40, while the object must be carried 2.9 m after the loss.The velocity estimate must hold across the whole carry.Source: G FAILURE CASES · p. 17
  • With thruster 1 delivering no thrust—a fault absent from play—goto succeeds in 4 of 40 trials versus U0’s 1 of 40.A weakened thruster is largely absorbed by the disturbance token, with 21 of 40 successes at half thrust; a missing one is not.Source: G FAILURE CASES · p. 17
  • In turbid water with a Jerlov parameter of 0.50, AquaWAM reaches the water tower in 9 of 20 trials versus U0’s 1 of 20.The water condition degrades the cameras that supply the goal.Source: G FAILURE CASES · p. 17

H QUALITATIVE COMPARISONS

Selected episodes illustrate task outcomes and duration differences, while the navigation, tracking, and transport panels do not share initial conditions.

H QUALITATIVE COMPARISONS

  • In shared episodes, AquaWAM grasps pick-blue objects successfully, and completes pipeline-pool inspection in 53 seconds versus 95 for π0.5 and 122 for U0.The figure also includes two grasps and an inspection episode from identical initial conditions.Source: H QUALITATIVE COMPARISONS · p. 18, Figure 8 · p. 18
  • On scan-ship-modern, GR00T N1.5, AquaWAM, and U0 finish in 61, 63, and 67 seconds, respectively, while X-VLA takes 173 seconds.Across the four shared episodes, π0.5 succeeds only on inspection, X-VLA only on scanning, and SmolVLA and OpenVLA on none.Source: H QUALITATIVE COMPARISONS · p. 18
  • Navigation, tracking, and transport panels show AquaWAM episodes, but lack shared initial conditions, so their durations are not directly comparable.The panels use the corresponding model’s episode when recorded, otherwise AquaWAM’s median-duration episode.Source: H QUALITATIVE COMPARISONS · p. 18, Figure 9 · p. 19

Related papers

Loading 2609.33299v2…