World Action Models (WAMs) have emerged as a mainstream technical approach to embodied intelligence, offering notable advantages in generalization and environmental perception. However, their large parameter counts, substantial computation per inference, and high inference latency remain significant limitations.
As the team behind the world's first edge-native embodied AI brain, Motion Brain has accumulated five years of experience across seven generations of optimized models, spanning model acceleration, inference latency optimization, and hardware–software co-optimization for edge deployment. In recent months, the team has focused on deploying WAMs at the edge and enabling real-time interaction.
On October 14, 2026, researchers from Motion Brain and Fudan University jointly introduced 1Step-WAM, described as the world's first edge-deployable World Action Model. The architecture is designed to substantially reduce inference latency during robot action generation by transforming the multi-step denoising process of a world action model into single-step action generation at the model-architecture level.
To stack one cup on top of another, a robot must determine their positions, generate an action, and then align them with its robotic arm. Action accuracy determines whether the cups can be stacked securely, while the time required for the model to compute the action also affects how the robot subsequently interacts with its environment.
A World Action Model (WAM) combines visual world modeling with action generation, allowing a robot to generate actions based on predictions of environmental changes. In practice, however, producing an action sequence often still requires multiple rounds of computation. Even when the visual component has been accelerated, action-generation latency can become a new bottleneck.
1Step-WAM addresses the action-generation process itself. It uses lightweight calibration to compress multi-step denoising into a single forward pass, while overlapping visual and action computations. On the RoboTwin simulation benchmark, the method delivered end-to-end inference speedups of 5.98× to 24.84× across three World Action Model backbones, with average task success rates falling by less than one percentage point relative to their respective multi-step baselines.
Where Does the Time Go in Action Generation?
In one class of World Action Models, the visual branch first processes current observations and language instructions, extracting features for use by the action branch. This stage is known as visual prefill. The action branch then starts from noise and progressively produces an executable action sequence through multiple denoising steps.
Here, "denoising" can be understood as repeatedly refining an action sequence being generated. Each step requires a network computation. In a conventional sequential pipeline, the action branch must also wait until the entire visual branch has finished before it can begin.
Latency therefore comes from two sources: repeated computations within action generation itself, and waiting time between the visual and action branches. Reducing the number of denoising steps shortens the first component; changing how the two branches are scheduled can further reduce waiting time.
How Much Action Information Is Already Present in the First Step?
While analyzing the action-generation process, the team observed an interesting pattern.
With visual observations, language instructions, and robot states held fixed, the researchers independently sampled 10,000 sets of initial noise for FastWAM. After multi-step denoising, the empirical variance of the final actions was only 1.182 × 10⁻⁵. This indicates that, under these fixed conditions, different noise initializations led to highly similar final actions.
Further trajectory analysis showed that the predicted directions at adjacent denoising steps were also highly similar, yielding nearly straight generation trajectories. Under the low-variance Gaussian model assumed in the paper, the velocity prediction from the first step could already provide a good estimate of the principal direction toward the final action.
"Velocity" here refers to the direction and magnitude of change during the model's generative process, not the physical speed of the robotic arm. This raised a research question: could the information available in the first step be used to directly produce an action sequence close to the multi-step result?
Simply reducing denoising to one step is not sufficient. On RoboTwin, the average success rate of a naïve single-step FastWAM variant dropped to 68.72%, compared with 91.83% for the multi-step baseline. Single-step generation therefore requires error compensation to preserve task performance.
Correcting Errors with Lightweight Calibration
Building on this observation, the team froze the original World Action Model backbone and introduced a lightweight calibration head. It takes the intermediate features, initial noise, and predicted velocity from the first step as inputs, and learns to correct that velocity so that the action generated in a single step approximates the output of the multi-step teacher model as closely as possible.
Only the calibration head is updated during training. For each of the three backbones, the number of trainable parameters is approximately 13.4 million. This figure refers to the additional parameters that must be trained, not the total parameter count of the World Action Model.
After calibration, the method also applies a Discrete Cosine Transform (DCT) along the temporal dimension of continuous joint actions, retains the first K low-frequency modes, and reconstructs the action sequence. The gripper dimensions are left unchanged.
This processing step follows from an analysis of the action spectrum: most of the teacher actions' energy is concentrated at low frequencies, while in some weaker high-frequency components the prediction error is larger than the true action signal. Discarding those frequencies can reduce residual errors in single-step predictions. However, retaining only low-frequency components is not necessarily beneficial for every action; an appropriate cutoff must balance preservation of action details against error suppression.
Coordinating Visual and Action Computation
Single-step generation also changes how the model is scheduled at runtime.
Each layer of the action branch depends on features produced by the corresponding visual layer, but does not need to wait for every subsequent visual layer to finish. Consequently, once the features from a visual layer are ready, the corresponding action layer can begin computing while the visual branch continues processing the next layer.
1Step-WAM uses separate CUDA streams and event synchronization to achieve this layer-by-layer coordination. It preserves the computational dependencies required by each layer while overlapping the execution of the two branches as much as possible.
Ablation studies show that, on top of single-step action generation, joint execution further reduced end-to-end latency by 33.95% to 45.30% across the three backbones. This demonstrates that the overall speedup comes both from fewer computations for action generation and from less sequential waiting between visual and action processing.
Validation in Simulation and on Real Robotic Arms
The team applied 1Step-WAM to FastWAM, ImageWAM, and Kairos, and evaluated it on two simulation benchmarks, RoboTwin and LIBERO.
RoboTwin contains 50 dual-arm manipulation tasks covering standard and randomized environments. LIBERO provides complementary evaluation through four task suites focusing on spatial relationships, objects, goals, and long-horizon tasks.
| Model | Original Latency (ms) | Single-Step Latency (ms) | Speedup | Original Success Rate (%) | Single-Step Success Rate (%) |
|---|---|---|---|---|---|
| FastWAM | 669.54 | 89.95 | 7.44× | 91.83 | 90.89 |
| ImageWAM | 306.85 | 51.28 | 5.98× | 93.34 | 92.44 |
| Kairos | 1831.62 | 73.73 | 24.84× | 95.96 | 95.20 |
The figures above are RoboTwin results, with inference latency measured on a single NVIDIA L40 GPU. Average success rates across the three World Action Model backbones declined by less than one percentage point.
On LIBERO, 1Step-WAM likewise achieved average success rates close to those of the corresponding multi-step baselines. For Kairos, the single-step version of 1Step-WAM reached a 97.2% success rate, exceeding its multi-step baseline by 1.0 percentage point.
For real-robot experiments, the team used an AIRBOT Play robotic arm equipped with a fisheye camera to evaluate three tasks: pick-and-place, cup stacking, and cup hanging.
The FastWAM-based 1Step-WAM achieved a 7.44× inference speedup relative to the FastWAM results reported in the paper, with an average success rate of 75% across the three tasks. Each method was run independently 20 times per task.
Cup stacking and cup hanging place greater demands on precise alignment. On these two tasks, naïve single-step generation achieved success rates of 40% and 10%, respectively, while 1Step-WAM reached 80% and 45%, demonstrating a substantial improvement after calibration.
What Does Faster Action Generation Mean?
For embodied AI models, inference efficiency determines how quickly the system can obtain a new action sequence and how much time remains for subsequent perception and decision-making.
By using lightweight calibration to reduce repeated computation during multi-step denoising, and overlapping the execution of the visual and action branches to cut waiting time, 1Step-WAM enables World Action Models to generate the next action sequence more promptly.
As part of ongoing research into the Motion Brain general-purpose embodied AI architecture, the Fudan University and Motion Brain research teams developed 1Step-WAM to explore how task performance and inference efficiency can be maintained with relatively low additional training cost.
The work offers a new technical approach to responsive, continuous interaction in embodied AI systems.
A Global Leader in Edge-Deployable Embodied AI Models
As the first research team in the industry to propose edge-deployable embodied AI models, Motion Brain has accumulated seven generations of model development focused on lightweight deployment at the edge.
In 2022, the Motion Brain team, together with Fudan University and NVIDIA, introduced GF-SQ, described as the first multimodal model optimization framework for edge deployment. The paper, "Boost Vision Transformer with GPU-Friendly Sparsity and Quantization," was accepted at CVPR 2023. The framework can reduce the FLOPs required for inference in large multimodal models by a factor of 62.
In March 2023, the Motion Brain team introduced PAGCP, a performance-aware global channel pruning framework for multimodal models. The paper, "Performance-aware Approximation of Global Channel Pruning for Multitask CNNs," was published in IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) in 2023. The approach enables 3.3× faster inference on edge devices while reducing parameter counts by 78%, with largely comparable performance across multiple models.
In February 2024, the Motion Brain team proposed Multimodal Alignment-Guided Dynamic Token Pruning (MADTP), a new framework for accelerating multimodal models. The paper, "MultimodalAlignment-Guide Dynamic Token Pruning for Accelerating Vision-Language Transformer," appeared at CVPR 2024. MADTP substantially reduces computational complexity across multimodal models while maintaining competitive performance. For some multimodal models, it reduces GFLOPs by 80% with a performance decline of less than 4%.
In May 2025, the Motion Brain team proposed EAI-CB, described as the world's first edge-side compression framework for large embodied AI models. Its paper, "Boost Embodied AI Models with Robust Compression Boundary," was accepted at IJCAI 2025. With 75% of model parameters compressed, the framework delivered performance improvements of 3%–6.2% on certain tasks.
In November 2025, the EAI-CB work on edge-side compression of large embodied AI models was recognized as an IJCAI 2025 Distinguished Paper. According to the original article, the team was the only Chinese recipient in six years; out of more than 5,500 submissions to IJCAI 2025, only three papers received this distinction. The award reflects international recognition of the team's expertise in compressing embodied AI models for edge deployment.
In December 2025, the Motion Brain team introduced MADTP++, a new framework that unifies customized token pruning and weight pruning. Its paper, "MADTP++: Bridge the Gap between Token and Weight Pruning for Accelerating VLTs," appeared in IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) in 2025. The approach substantially reduces both model parameters and computational overhead while maintaining competitive performance; some models even recorded a 2% performance improvement despite an 80% reduction in parameters.
Looking ahead, Motion Brain will continue exploring performance optimization for embodied AI brains deployed at the edge, advancing the development of edge-deployable embodied AI models worldwide and improving responsiveness and operational stability in real-world robotic tasks.
