MotionWAM: Towards Foundation World Action Models for Real-Time Humanoid Loco-Manipulation
1. Paper overview
In one sentence: MotionWAM turns a single pass through a video world model into whole-body action features, improving G1 task success while leaving its precise motion-token encoding unresolved. motivationvideo-featuresmotion-interfacetraining-stagesevaluationstage-ablationefficiency
| At a glance | What to know |
|---|---|
| Research problem | Source description Humanoid tasks such as kicking a ball or retrieving a basket under a table require task-directed leg and torso movement. The authors identify two obstacles: interfaces that restrict legs to coarse locomotion commands, and video-action denoising too expensive for feedback control. They propose a unified motion vocabulary plus inexpensive access to a video dynamics prior. motivationevaluation |
| Core mechanism | Source description A separately parameterized Video DiT and Motion DiT exchange hidden features, while SONIC converts predicted whole-body intent into executable joint commands. Continuous end-effector channels accompany the motion representation. video-featuresmotion-interface |
| A key reported result | Nine real-world Unitree G1 loco-manipulation tasks: 76.1% Mean task success rate (%). 20 physical trials per task per method; identical Stage 3 demonstrations, observations and unified SONIC action interface. GR00T-N1.7: 43.9%, the strongest tested baseline. A 32.2-percentage-point gain using reported rounded means. MotionWAM leads every task, but this is a familiar-object task suite, with no reported confidence intervals. evaluationtask-resultslimits |
| Reading caution | Source description Validation is confined to the Unitree G1. No controlled novel-object study is reported; train/test objects share visual similarity. The authors attribute failures to targets leaving the egocentric field of view or camera viewpoints drifting from training conditions, causing stalls or inaccurate trajectories. limits |
Core contributions
- Source description
A separately parameterized Video DiT and Motion DiT exchange hidden features, while SONIC converts predicted whole-body intent into executable joint commands. Continuous end-effector channels accompany the motion representation. video-featuresmotion-interface
- Author claim
The authors present the system as the first real-time, closed-loop WAM for whole-body humanoid loco-manipulation with task-driven foot interaction. This priority claim is not independently established by the supplied comparison. motivation
Figure 3. Video adaptation precedes action grounding and task-specific whole-body fine-tuning. Original paper, p. 5 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Begin with the purple Stage 1 panel: robot and human egocentric videos adapt the Video DiT. The gray copy arrow transfers that trunk into Stage 2; it denotes weight transfer, not a robot action. Embodiment tags route heterogeneous action interfaces around the Motion DiT. Section 3.2 supplies the coupling omitted from this schematic: the motion branch attends to hidden states from one video forward pass. Along the bottom, SMPL24 retargeting and a controller produce teleoperation demonstrations for Stage 3. The rightmost output combines a discrete motion representation with continuous end-effector values, then drives the robot through the controller described in the text. video-featurestraining-stagesmotion-interfaceteleoperationconfiguration
What it supports. The figure organizes two different transfers: video knowledge enters the action model, and human teleoperation supplies target-robot behavior. The VAE and language encoder remain frozen throughout; only the video trunk trains in Stage 1, while both DiT trunks train in Stages 2 and 3.
Where the evidence stops. The output graphic's '64-dim codebook' and positions 0–63 do not resolve the text's 64-dimensional vector versus scalar-index descriptions. Eq. (6) specifies scalar rounding, whereas Table 4 lists 66 action channels. Their mapping remains unspecified.
2. Motivation
2.1 The problem and the proposed response
Humanoid tasks such as kicking a ball or retrieving a basket under a table require task-directed leg and torso movement. The authors identify two obstacles: interfaces that restrict legs to coarse locomotion commands, and video-action denoising too expensive for feedback control. They propose a unified motion vocabulary plus inexpensive access to a video dynamics prior. motivationevaluation
2.2 What this reading follows
A humanoid lifting a basket from under a table must coordinate viewpoint, height, feet and hands. MotionWAM connects that coordination problem to video pretraining: a video transformer provides hidden features to a separate motion transformer, which predicts commands through a shared whole-body interface. The system learns visual dynamics first, action grounding second and target tasks last. Its physical-robot evaluation is encouraging, but interpreting the gains requires separating backbone choice, training exposure and action representation. Read the architecture together with the ablation and timing tables: the policy avoids complete future-video generation, yet its reported speed counts action chunks and still depends on a low-level controller. motivationvideo-featuresmotion-interfacetraining-stagesevaluationstage-ablationefficiency
3. Research context
We place the paper in the collection through its world–action interface. The catalog labels and the reading’s assessment are shown separately.
| Catalog dimension | Recorded classification |
|---|---|
| Major category | WAMs |
| Architecture | Dual-system |
| Prediction paradigm | IDM |
| Quadrant | Q4 · Dual-system × IDM |
3.1 Evidence-based assessment
Supports the recorded classification
The core Dual-system × IDM assignment is supported by two distinct DiT trunks and the explicit predict-video-dynamics–then-invert formulation. Joint flow training does not turn them into one shared backbone. Here inversion uses intermediate video features, not a fully generated future or planning search. The catalog's generalization subcategory should not be read as evidence of controlled novel-object or cross-platform transfer. video-featuresflow-losseslimits
This is the collection’s architectural analysis, not a new related-work survey. Benchmark comparisons and their protocols appear in Section 6.
4. Problem formulation
4.1 Inputs and outputs
| Inputs | Outputs |
|---|---|
|
|
4.2 Equations and their role
5. Method in detail
5.1 Start with the action vocabulary, then identify the controller
Imagine asking the G1 to lift a basket from under a table. The task prompt requires moving the basket onto the table, while the illustrated trajectory involves lowering and repositioning the body. MotionWAM predicts whole-body intent instead of providing detailed upper-body targets alongside coarse lower-body commands. SONIC remains the execution mechanism, and continuous end-effector channels provide hand control. Reader interpretation: the useful change is what the high-level action space can express; calling it unified does not remove the low-level controller. For implementation, retain the source's distinction between motion representation and joint commands. Section 3.1 describes a discretized vector, Figure 3 shows a codebook-like output, and Eq. (6) describes scalar rounding. Those descriptions establish the intended information flow but do not supply a consistent token mapping. evaluationrolloutsmotion-interfaceteleoperationmotivation
Table 4. The compact deployment path sits on top of substantial staged training. Original paper, p. 14 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the table in three blocks. The video block names Cosmos-Predict2.5-2B, 2048-dimensional features and one future flow inference step; Section 3.2 clarifies that the action policy reads hidden states rather than finishing future-frame denoising. The motion block gives four inference timesteps, 66 action channels and 64 state channels. The lower block separates stage-dependent budgets from shared settings: GPU counts fall from 128 to 32 to 8, while maximum training steps fall from 100,000 to 50,000 to 15,000. Batch size remains eight per device. The frozen-module row applies across stages, including the final fine-tune. configurationvideo-featuresmotion-interfacetraining-stages
What it supports. Training scale and inference cost answer different questions. One video pass and four motion timesteps describe the deployment computation, while the stage columns reveal the adaptation budget required before deployment. This table supports a staged training recipe rather than a claim that the complete system is cheap to reproduce.
Where the evidence stops. Training GPU models and deployed chunk horizons are not specified here. The conditional-frame timestep is distinct from the noisy-future hook timestep in Eq. (2); maximum sequence length is not a declared action-chunk length. The token dimensional mismatch also remains.
5.2 Use a future-prediction objective without rendering the future
The video objective trains a velocity field between clean future latents and Gaussian noise. That gives the Video DiT a predictive training task, but deployment reads a different product of the network: hidden activations from a single transformer block. In Eq. (2), the observed frames condition a pass whose future input remains near the pure-noise endpoint. The Motion DiT consumes the resulting features together with proprioception, then performs its own motion generation. Table 4 lists four motion inference timesteps. Thus one-shot video feature extraction does not imply that the whole policy is one feed-forward operation. The source's predict-then-invert terminology refers here to feature-conditioned action prediction, not to first decoding a finished movie. During Stages 2 and 3, the video loss continues alongside the motion loss to preserve predictive representation learning. video-featuresflow-lossesconfigurationtraining-stages
5.3 Ask which comparison supports which mechanism
The headline result compares complete systems after a shared target-task fine-tune. It shows that MotionWAM succeeds more often across the nine-task suite, including kicking, but every tested baseline already emits unified motion latents. Reader analysis: that experiment cannot assign the gain to the action-space change. Table 1 asks a narrower question by removing video adaptation or action post-training while preserving Stage 3. Its losses support both stages under the chosen training budgets, without proving that their order is uniquely optimal. Qwen3DiT retains the motion trunk and downstream stages, yet changes the backbone and omits Stage 1; it is a useful comparison with a remaining training-history difference. Finally, Table 2 measures chunk throughput on an A100. It complements the task evaluation but cannot supply the unreported RTX 4090 latency. task-resultsevaluationbaseline-controlsstage-ablationefficiencyteleoperation
5.4 Training and inference
During training
Stage 1 updates only the Video DiT on approximately 2,136 hours of video, ignoring action labels. Table 5 allocates 30% to human video, 50% to the domain labeled G1-class humanoids and 20% to other robots; the listed sources also include simulated Fourier GR1 video. Within-domain source weights follow the square root of episode count. training-stagesdata-mixture
Stage 2 jointly updates video and motion trunks on heterogeneous action-labeled humanoid data. Actions are right-padded to 66 channels with validity masks. The video flow loss remains as a representation regularizer alongside the motion loss. Stage 3 retains both losses and adapts to 200 demonstrations for each of nine G1 tasks, recorded at 50 Hz. VAE and text encoder stay frozen throughout. training-stagesflow-lossesdata-mixtureconfiguration
Table 4 specifies 100,000/50,000/15,000 maximum steps and 128/32/8 GPUs for Stages 1/2/3, with batch size eight per device. Video learning rate is 10⁻⁵ throughout; motion learning rate is 10⁻⁴ in Stages 2–3. AdamW and a cosine schedule are specified, but these rows do not identify MotionWAM training GPU models. configuration
During inference
Each policy query obtains video hidden states in one pass, without completing future-frame denoising. The motion branch uses four inference timesteps, then supplies motion and end-effector outputs to the controller. The onboard system queries a WebSocket policy server on an RTX 4090 workstation in closed loop; the separate A100 frequency experiment measures action-chunk emission. video-featuresconfigurationteleoperationefficiency
5.5 Implementation flow
- Extract predictive visual features
A causal spatiotemporal VAE encodes frames. The Video DiT starts from Cosmos-Predict2.5-2B and uses Cosmos-Reason1 language embeddings. A hook reads one transformer block at a fixed future flow time near pure noise, with clean observed-frame latents as conditioning. video-features
- Generate embodiment-conditioned motion
The Motion DiT combines hooked video features, proprioception and noisy motion tokens through interleaved self/cross-attention. Embodiment-specific input/output projectors surround a shared motion trunk; deployment uses the Unitree G1 projector. video-featurestraining-stages
- Decode whole-body intent
SONIC supplies the whole-body controller, while gripper or dexterous-hand channels remain continuous. Section 3.3 describes regressing a scalar motion-token index with flow matching and rounding it before decoding; the conflicting dimensional descriptions remain unresolved below. motion-interface
6. Experiments & results
MotionWAM conditions a whole-body motion generator on features from one forward pass of a video world model. Its three-stage training recipe leads to 76.1% mean success on nine physical Unitree G1 tasks, versus 43.9% for the strongest tested baseline. The key tradeoff is access to video-trained representations without fully generating future frames; token-interface ambiguities and limited generalization tests constrain reproduction and interpretation.
6.1 Read the original evidence
Figure 5. MotionWAM leads the tested policies on every task under the common action interface. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Use the legend to compare pink MotionWAM with green GR00T-N1.7 before checking the other four policies. Each bar is a percentage from 20 physical trials, so one trial changes a task score by five percentage points. Kick Soccer tests foot interaction; Retrieve Item and Lift Basket combine body movement with object handling. Wipe Board reaches 95% for MotionWAM and 50% for GR00T, while Toss Garbage remains comparatively difficult at 45% versus 30%. The paper gives nine-task means in the adjoining text, not as an extra bar group. All policies use the same Stage 3 demonstrations and unified SONIC interface. task-resultsevaluationbaseline-controlsrollouts
What it supports. The reported mean rises from 43.9% for GR00T-N1.7 to 76.1% for MotionWAM, a 32.2-percentage-point increase. Kick Soccer improves from 20% to 60%. These are executed task-success measurements, distinct from video-generation quality or the qualitative rollout montage.
Where the evidence stops. The chart supplies no confidence intervals or repeated-seed statistics. Shared action outputs mean it does not test a unified-versus-split body interface. Qwen3DiT also changes backbone and omits Stage 1, limiting isolation of the video prior.
Table 2. Skipping complete video denoising improves WAM throughput, measured per action chunk. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read parameter count and frequency together: these policies are of comparable scale but are not parameter-identical. Section 4.4 and the original caption specify a single NVIDIA A100 and define frequency as emission of a whole action chunk. MotionWAM's 4.9 Hz is seven times Cosmos Policy's 0.7 Hz. The source attributes that difference to reading intermediate video features in a single pass instead of iteratively denoising future video. The two VLA-style alternatives remain faster. Keep this benchmark separate from the real-robot setup in Sections 4.1 and B, which uses an RTX 4090 policy server and an onboard controller. efficiencyvideo-featuresconfigurationteleoperation
What it supports. The measured result establishes a substantial throughput improvement over the listed video-denoising comparator. It does not establish the fastest policy overall: GR00T-N1.7 reaches 6.5 Hz and Qwen3DiT reaches 9.0 Hz. The evidence concerns chunk production, while SONIC handles whole-body command execution.
Where the evidence stops. The 4.9 Hz figure is neither a motor servo rate nor a measured RTX 4090 result. The paper leaves exact timing boundaries and MotionWAM chunk horizon unspecified, preventing a complete end-to-end latency reconstruction.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| Nine real-world Unitree G1 loco-manipulation tasks 20 physical trials per task per method; identical Stage 3 demonstrations, observations and unified SONIC action interface. | 76.1% Mean task success rate (%) | GR00T-N1.7: 43.9%, the strongest tested baseline. A 32.2-percentage-point gain using reported rounded means. MotionWAM leads every task, but this is a familiar-object task suite, with no reported confidence intervals. evaluationtask-resultslimits |
| Kick Soccer Kick the ball into the goal; 20 real-world G1 trials under the shared evaluation protocol. | 60% Task success rate (%) | GR00T-N1.7: 20%; π0.5: 10%; Qwen3DiT, Diffusion Policy and ACT: 0%. Executed foot interaction improves by 40 percentage points over GR00T. Because every tested policy has unified motion outputs, this does not isolate the benefit of replacing a decoupled action interface. evaluationtask-resultsbaseline-controls |
| Training-stage ablation on five representative tasks Lift Basket, Retrieve Item, Load Cart, Toss Garbage and Kick Soccer; 20 trials per task; Stage 3 retained in every variant. | Full: 70.0%; without Stage 1: 59.0%; without Stage 2: 42.0%. Mean task success rate (%) | Removing Stage 1 costs 11 percentage points; removing Stage 2 costs 28. Both preparatory stages help in this setup. This five-task mean must not be compared directly with the nine-task 76.1%; stage removal also removes training exposure. stage-ablation |
| Policy deployment frequency Single NVIDIA A100; frequency counts complete action chunks emitted. | 4.9 Hz at 2.5B trainable parameters Chunk emission frequency (Hz) | Cosmos Policy: 0.7 Hz, 2.0B; GR00T-N1.7: 6.5 Hz, 1.6B; Qwen3DiT: 9.0 Hz, 2.3B. Seven times Cosmos Policy's reported frequency, while slower than both listed VLA-style alternatives. This is neither the joint servo rate nor a measured RTX 4090 rate. efficiencyteleoperation |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Table 1. Both preparatory stages contribute before the shared final fine-tune. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Check the Stage 1 and Stage 2 columns before comparing task scores. Every row still includes Stage 3; a dash removes only the indicated earlier stage. Read the Full row as the reference, then move upward to see what is lost. Removing video adaptation changes the average from 70.0 to 59.0, while removing action post-training changes it to 42.0. Retrieve Item is especially sensitive to Stage 2 removal: 90 becomes 45. The five columns cover Lift Basket, Retrieve Item, Load Cart, Toss Garbage and Kick Soccer, so this average uses a smaller task set than the main result. stage-ablationtraining-stages
What it supports. The larger drop after removing Stage 2 supports the value of action grounding before target-task fine-tuning, while the Stage 1 drop supports egocentric video adaptation. Both conclusions apply to this five-task, 20-trial-per-task evaluation with the final fine-tune retained.
Where the evidence stops. Removing a stage also removes its training data and optimization budget. The table has no equal-compute replacement, Stage 3 removal, or from-scratch full-training control; it therefore does not establish that this ordering is uniquely necessary.
Figure 6. A whole-body motion vocabulary cannot compensate for losing the task object from view. Original paper, p. 16 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read each row across: the left image shows the robot and task context from outside, while the right shows what the policy's head-mounted camera sees. In Lift Basket, the external view reveals the basket beside the table and the lowered robot; the ego image is dominated by the robot's arm and floor. In Kick Soccer, the external camera shows the ball and goal, while the ego image mostly shows floor and the robot itself. These pairs illustrate the information available at a failure, rather than a time series showing exactly when the object disappeared. limits
What it supports. Appendix F attributes stalls or inaccurate whole-body trajectories to lost visual grounding when an object exits the camera view or the viewpoint departs from the training distribution. The figure makes that proposed failure mechanism concrete for both manipulation near the floor and task-directed foot interaction.
Where the evidence stops. These are selected diagnostic examples, not an occlusion ablation or a measured distribution of failure causes. They do not establish causality, recovery probability, or novel-object generalization, which the limitations section explicitly leaves untested.
7. Analysis & limitations
7.1 What the evidence leaves open
Validation is confined to the Unitree G1. No controlled novel-object study is reported; train/test objects share visual similarity. The authors attribute failures to targets leaving the egocentric field of view or camera viewpoints drifting from training conditions, causing stalls or inaccurate trajectories. limits
The interface specification is inconsistent: Section 3.1 describes a 64-dimensional discretized SONIC vector and '2 tokens of 32 levels'; Figure 3 labels a '64-dim codebook' with positions 0–63; Section 3.3 describes one scalar token index; Table 4 gives 66 action channels. The source does not reconcile these representations or specify the index-to-vector mapping. motion-interface
Common demonstrations and action interfaces strengthen the comparison but do not make the video prior the sole changed factor. Qwen3DiT keeps the motion trunk and Stages 2–3 but changes the backbone and omits Stage 1; other baselines use different pretraining and recipes. No isolated unified-versus-split action-space ablation is reported. baseline-controlsstage-ablation
7.2 Questions for discussion
- Would motion prediction still improve when the video and static-backbone alternatives receive matched egocentric adaptation?
- How should the scalar index in Eq. (6) map to the 64-dimensional SONIC representation described earlier?
- Does restoring target visibility recover performance under controlled viewpoint shifts?
8. Reproducibility audit
8.1 Requirements and known gaps
The demonstrated stack requires G1, dual ALOHA 2 grippers, a head-mounted D435i, SONIC and the policy server. Collection uses PICO tracking through XRoboToolkit, a 24-joint SMPL pose retargeted to 29-DoF G1 commands, and LeRobot episodes at 50 Hz. Appendix B mentions ankle trackers and hand controllers beyond the main text's abbreviated 'three-point' description. teleoperation
Resolve the token mapping before implementation. Also missing are the selected video hook block, concrete future-video and motion-chunk horizons, exact Stage 2 source weights/counts, and fully specified timing boundaries. Table 5 details Stage 1 only; Table 4's maximum sequence length is not an explicit deployed chunk horizon. Official recipes are referenced for two VLA baselines without complete local settings. video-featuresmotion-interfaceconfigurationdata-mixtureefficiencybaseline-controls
Proposed checks should separate predictive features from backbone/training exposure and test the stated visual-grounding failure mechanism. Match motion interfaces, data and action timing; measure executed task success as well as feature or latency changes. These are proposed experiments, not reproduced findings. video-featuresbaseline-controlslimits
8.2 Proposed reproduction checks
The following checks are proposals motivated by the paper. They have not been run as part of this reading.
Check 1: Test whether temporal adaptation, rather than extra exposure, explains Stage 1
Reader-proposed experiment: start two copies of the same Cosmos checkpoint and use identical Stage 1 videos, frame counts, optimizer budgets and sampling weights. Preserve temporal order for one; shuffle future-frame order within clips for the control while preserving image exposure. Apply identical Stages 2 and 3, Motion DiT interfaces and inference timesteps. Evaluate the five Table 1 tasks over repeated training seeds and report task success with uncertainty. A consistent advantage for ordered-video adaptation would support temporal learning beyond exposure; parity would weaken that explanation. This is a proposed diagnostic, with sequence corruption itself an acknowledged distribution change. video-featurestraining-stagesdata-mixtureconfigurationstage-ablationbaseline-controls
Check 2: Separate loss of target visibility from difficulty of the body motion
Reader-proposed experiment: repeat Lift Basket and Kick Soccer from matched object and robot starting poses using the same policy. Compare a sham camera intervention, brief target occlusion with viewpoint otherwise fixed, and a controlled camera shift; then restore the original view. Keep action timing fixed and log target visibility, stalls, task success and time to resume useful motion. If performance degrades specifically during lost visibility and recovers when the target returns, the Appendix F explanation gains support. Persistent failures with a visible target would indicate additional control or distribution-shift causes. No such intervention study was reported or run here. limitsevaluationteleoperationefficiency
8.3 Reading coverage
Visual audit: All five supplied text chunks were read completely, including references and Appendices A–F. Original page images were inspected for the title/author/version block, Figures 1–7, Tables 1–5, Eqs. (1)–(6), training, evaluation, hardware, data and failure details. The six final original crops were individually viewed and retain their labels, legends and table boundaries; no charts were recreated. Figure 3's arrows and output markers were checked against Sections 3.1–3.3, Eq. (6) and Table 4; its unresolved token representation is disclosed. Pages 9–12 contain references and were read in text only. Separate supplements, project videos, code and datasets remain outside this reading. All pages supporting the retained technical claims and proposed checks are included above.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 13, 14, 15, 16, 17. Appendix coverage: reviewed.
Original figures and tables remain the work of the source’s authors. Extractions preserve their scientific content; any HTML wrapper layout is disclosed with each figure. The surrounding reading notes are our own.
Text reading scope & known omissions
- Title, author block and abstract (pp. 1–2)
- 1 Introduction (pp. 2–3)
- 2 Related Work (pp. 3–4)
- 3 Method, including 3.1–3.3 and Eqs. (1)–(6) (pp. 4–6)
- 4 Experiments, including 4.1–4.4 (pp. 6–8)
- 5 Conclusion and 6 Limitations (p. 8)
- References (pp. 9–12)
- A Real-World Task Suite and A.1 Per-Task Language Prompts (p. 13)
- B Whole-Body Teleoperation Setup, B.1–B.2 (p. 13)
- C Model and Training Configuration (pp. 13–14)
- D Baseline Training Configurations (pp. 13–15)
- E Per-Stage Data Composition (pp. 15–16)
- F Failure Cases and final inference demonstrations (pp. 16–17)
Outside the original text pass
- Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
- Separate supplemental material availability has not been fully verified.
- The extraction's image limitation was addressed by inspecting original PDF pages 1–8 and 13–17 and all seven figures and five tables. Reference pages 9–12 were read as text.
- Identity/version: the inspected title page identifies arXiv:2606.09215v1 [cs.RO], 8 June 2026. Title and all six ordered authors match the catalog after name-order normalization; no title or author revision discrepancy was observed. No other edition was supplied or compared.
- No separate supplement was supplied. Project pages, code, checkpoints and datasets were not inspected; experiments were not reproduced.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
identityPDF p. 1, title/author block and arXiv margin
Title: MotionWAM: Towards Foundation World Action Models for Real-Time Humanoid Loco-Manipulation. Authors in order: Jia Zheng, Teli Ma, Yudong Fan, Zifan Wang, Shuo Yang, Junwei Liang. Affiliations: Mondo Robotics; HKUST (GZ); HKUST. The margin identifies arXiv:2606.09215v1 [cs.RO], 8 Jun 2026.
Go to primary source ↓motivationPDF pp. 1–3, abstract, Section 1, Figure 2 and contributions
The authors motivate unified whole-body tokens and faster video-conditioned action prediction; the stated priority claim concerns real-time closed-loop whole-body WAM deployment.
Go to primary source ↓video-featuresPDF p. 4, Sections 3.1–3.2, Eqs. (1)–(2)
Two DiT trunks implement video-feature-conditioned motion prediction. A single hooked video pass at τ_f approximately 1 consumes observed latents and noisy future latents; the motion branch uses video features, proprioception and embodiment projectors.
Go to primary source ↓motion-interfacePDF p. 4, Section 3.1 whole-body latent; p. 5, Figure 3 output block/caption; p. 6, Section 3.3 and Eq. (6); p. 14, Table 4 action dimension; p. 15, ACT paragraph
The text alternates between a 64-dimensional discretized SONIC vector and a scalar token index rounded before decoding. Figure 3 labels a 64-dim codebook with positions 0–63; Table 4 and the ACT adaptation specify 66 action channels. Continuous hand/end-effector channels accompany whole-body intent.
Go to primary source ↓flow-lossesPDF p. 5, Section 3.3, Eqs. (3)–(4); p. 6, Eq. (5)
Video and motion objectives regress Gaussian-noise-minus-clean velocity targets. Stage 2 sums both losses; the adjoining Stage 3 description retains the sum.
Go to primary source ↓training-stagesPDF pp. 5–6, Figure 3 and Section 3.3, Stage 1–3 paragraphs
Stage 1 trains the video trunk on approximately 2,136 hours; Stage 2 attaches and co-trains the motion trunk across heterogeneous humanoid data; Stage 3 uses 200 episodes per target task. VAE and text encoder remain frozen.
Go to primary source ↓data-mixturePDF pp. 15–16, Appendix E and Table 5
Stage 1 lists normalized source weights, human/humanoid/other-robot budgets of 30/50/20%, and square-root episode weighting. Fourier GR1 simulation is one listed source. Stage 2 describes padded 66-channel actions and validity masks without an equivalent mixture table; Stage 3 gives 200 episodes per task at 50 Hz with one embodiment tag.
Go to primary source ↓teleoperationPDF p. 6, Section 4.1 Hardware platform; p. 13, Appendix B.1–B.2
G1 uses dual ALOHA 2 grippers and head-mounted D435i RGB; the WebSocket policy server runs on a single RTX 4090. Appendix B lists PICO headset, two ankle trackers and hand controllers, XRoboToolkit, SMPL-24 to 29-DoF G1 retargeting with SONIC, and 50-Hz LeRobot recording.
Go to primary source ↓evaluationPDF pp. 6–7, Section 4.1 and Figure 5 caption; p. 13, Appendix A.1, Table 3
Nine real-world tasks use 20 trials per task and common Stage 3 demonstrations. Table 3 specifies the prompts, including kicking into the goal and moving a basket from under a table onto it.
Go to primary source ↓task-resultsPDF p. 7, Figure 5 bars/legend and Section 4.2
MotionWAM/GR00T-N1.7 success percentages by task are PnP Bottle 95/80, Kick Soccer 60/20, Retrieve Item 90/50, Load Cart 75/35, Toss Garbage 45/30, Lift Basket 80/45, Stock Shelf 70/40, Wipe Board 95/50, Do Laundry 75/45. Text reports means 76.1/43.9. Kick Soccer is 10 for π0.5 and 0 for the remaining baselines. No uncertainty bars or intervals are supplied.
Go to primary source ↓stage-ablationPDF pp. 7–8, Section 4.3; p. 8, Table 1, all rows
Stage 3 is retained. In Lift Basket/Retrieve Item/Load Cart/Toss Garbage/Kick Soccer order, full scores are 80/90/75/45/60 (mean 70), without Stage 1 70/75/60/35/55 (59), without Stage 2 65/45/30/30/40 (42); each task has 20 trials.
Go to primary source ↓efficiencyPDF p. 8, Section 4.4 and Table 2, all rows/caption
The single-A100 benchmark measures action-chunk frequency: MotionWAM 2.5B/4.9 Hz; Cosmos Policy 2.0B/0.7 Hz; GR00T-N1.7 1.6B/6.5 Hz; Qwen3DiT 2.3B/9.0 Hz. The authors attribute the WAM speed difference to avoiding iterative future-video denoising.
Go to primary source ↓limitsPDF p. 8, Section 6; p. 16, Appendix F and Figure 6
The authors state G1-only validation, no controlled novel-object generalization, similar training/test objects, and loss of visual grounding under target invisibility or viewpoint drift. Figure 6 pairs third-person and ego views for Lift Basket and Kick Soccer, without failure-frequency statistics.
Go to primary source ↓configurationPDF p. 14, Table 4, Video DiT/Motion DiT/Training blocks
Video hidden features have dimension 2048; the motion hidden size/output dimension row is 2560. The table lists one future flow inference step, four motion inference timesteps, 66 action/64 state dimensions, maximum sequence length 1024, and flash-attention 2. Stage GPU counts are 128/32/8, maximum steps 100000/50000/15000, and per-device batch sizes 8/8/8; VAE/text stay frozen. Learning rates and AdamW/cosine settings are given, without MotionWAM training GPU models.
Go to primary source ↓baseline-controlsPDF p. 7, Sections 4.1–4.2; pp. 13–15, Appendix D, especially Qwen3DiT paragraph on p. 15
All policies share Stage 3 data and unified motion outputs. Qwen3DiT substitutes Qwen3-VL 2B while retaining the Motion DiT and Stages 2–3 but omitting video Stage 1. DP/ACT have separate architectures and recipes; GR00T-N1.7 and π0.5 are said to follow official fine-tuning recipes.
Go to primary source ↓rolloutsPDF p. 1, Figure 1; p. 6, Figure 4; p. 17, Figure 7
The overview, task grid and final rollout montage depict whole-body task behaviors. These are illustrative trajectories, separate from the success-rate evaluation.
Go to primary source ↓8.5 Primary sources
MotionWAM: Towards Foundation World Action Models for Real-Time Humanoid Loco-Manipulation ↗
PDF · 7,528 extracted words
Source fingerprint
7f1f828566788c1552b9a989b78e0ed525216d58218c2ac2247847e172f3c041