Paper deep dive
Rolling-WAM: World Action Models with Rolling Imagination
Yinghua Zhou, Junjie Ye, Yiqi Zhao, Hao Dong, Celina Shiyu Wang, Ruohai Ge, Tingyi Yang, Basile Van Hoorick, Gaurav Sukhatme, Vitor Guizilini, Yue Wang
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 94%
Last extracted: 9/26/2026, 3:08:11 AM
Summary
The paper introduces Rolling-WAM, a method for robotic manipulation that accelerates World Action Models (WAMs) by distributing joint video-action denoising across successive replanning cycles using a sliding window. This approach reduces inference latency by 4.5x while maintaining competitive success rates on benchmarks like LIBERO, RoboTwin, and real-world Unitree G1 tasks.
Entities (8)
Relation Signals (5)
Rolling-WAM → evaluatedon → LIBERO
confidence 95% · Evaluations on LIBERO... show that Rolling-WAM achieves competitive manipulation performance.
Rolling-WAM → evaluatedon → Robotwin
confidence 95% · Evaluations on... RoboTwin... show that Rolling-WAM achieves competitive manipulation performance.
Rolling-WAM → evaluatedon → Unitree G1
confidence 95% · Evaluations on... a real-world Unitree G1 humanoid show that Rolling-WAM achieves competitive manipulation performance.
Rolling-WAM → improves → Joint-WAM
confidence 95% · delivers a 4.5x steady-state replanning speedup over standard joint WAMs.
Rolling-WAM → uses → Wan2.2-TI2V-5B
confidence 90% · We initialize the video expert from Wan2.2-TI2V-5B
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:World Action Models (WAMs) couple action generation with future visual prediction for robotic manipulation. However, completing the joint video-action denoising process at each replanning cycle incurs substantial latency, delaying action updates and limiting closed-loop responsiveness. We present Rolling-WAM, a formulation that distributes joint denoising across successive replanning cycles. Our method maintains a sliding window of video-action chunks at staggered noise levels. At each step, a rolling noise schedule fully denoises the imminent action chunk for execution, while partially refining farther-future chunks. As the window advances with new camera observations, the retained future chunks continue their denoising process. This distributes the computational cost over time while carrying an evolving visual-action context across chunk boundaries. Evaluations on LIBERO, RoboTwin, and a real-world Unitree G1 humanoid show that Rolling-WAM achieves competitive manipulation performance. By removing the need to denoise the entire prediction horizon from scratch, it delivers a 4.5x steady-state replanning speedup over standard joint WAMs.
Tags
Links
- Source: https://arxiv.org/abs/2609.30247v1
- Canonical: https://arxiv.org/abs/2609.30247v1
Trouble viewing inline? Open PDF directly →
Full Text
47,862 characters extracted from source content.
Expand or collapse full text
Rolling-WAM: World Action Models with Rolling Imagination Yinghua Zhou1,2,* Junjie Ye1,* Yiqi Zhao1 Hao Dong1 Celina Shiyu Wang1 Ruohai Ge1 Tingyi Yang1,3 Basile Van Hoorick4 Gaurav Sukhatme1 Vitor Guizilini4,† Yue Wang1,† Affiliation: 1University of Southern California 2Brown University 3Fudan University 4Toyota Research Institute *Equal contribution †Equal advising Project page: https://rolling-wam.github.io/ Abstract World Action Models (WAMs) couple action generation with future visual prediction for robotic manipulation. However, completing the joint video-action denoising process at each replanning cycle incurs substantial latency, delaying action updates and limiting closed-loop responsiveness. We present Rolling-WAM, a formulation that distributes joint denoising across successive replanning cycles. Our method maintains a sliding window of video-action chunks at staggered noise levels. At each step, a rolling noise schedule fully denoises the imminent action chunk for execution, while partially refining farther-future chunks. As the window advances with new camera observations, the retained future chunks continue their denoising process. This distributes the computational cost over time while carrying an evolving visual-action context across chunk boundaries. Evaluations on LIBERO, RoboTwin, and a real-world Unitree G1 humanoid show that Rolling-WAM achieves competitive manipulation performance. By removing the need to denoise the entire prediction horizon from scratch, it delivers a 4.5×4.5× steady-state replanning speedup over standard joint WAMs. †aftertitle: Fig. 1: Rolling-WAM maintains a rolling window of video-action predictions at staggered noise levels. For simplicity, each frame represents one chunk, and the traces denote actions. After executing the first action chunk, the robot acquires a new camera observation, shifts the window, and denoises the retained future chunks together with a newly appended chunk initialized with Gaussian noise. This approach distributes the denoising cost across control cycles, achieving competitive task performance with substantially faster inference. I Introduction World Action Models (WAMs) have recently advanced robotic manipulation by jointly predicting robot actions and future observations [1, 2, 3]. This joint modeling provides visual context for action generation, helping the policy anticipate how the scene will evolve. However, deployment in dynamic environments requires closed-loop control: the robot must repeatedly replan from the latest observations to correct execution errors and respond to environmental changes. Frequent replanning exposes a computational bottleneck in standard WAMs. Joint denoising samplers process the entire prediction horizon from pure noise to clean data at each replanning cycle. Each denoising step processes the entire video-action sequence, making repeated model evaluations computationally expensive. The resulting inference latency delays action updates, limiting closed-loop responsiveness. Some methods achieve faster inference by omitting future-video generation at test time [4], at the cost of losing explicit visual predictions that could inform action generation. We argue that denoising can be organized more efficiently across replanning cycles. Conventional chunk-based samplers [2, 4] fully denoise a new prediction horizon at each replan, even when receding-horizon control executes only an initial portion. The unexecuted tail can inform current actions but is discarded rather than reused across replans. Retaining and progressively refining future video–action chunks across cycles can preserve predictive context and reuse denoising computation, reducing the sequential steps needed to produce the next executable chunk. We therefore introduce Rolling-WAM, a formulation that inherits this intuition by distributing denoising across successive replanning cycles. Building on the concept of rolling diffusion for sequence generation [5, 6], Rolling-WAM maintains a sliding window of aligned video-action chunks with progressively higher noise levels toward the future (Fig. ). At each replanning cycle, only the imminent action chunk is fully denoised for execution, while farther-future chunks remain partially denoised. This resembles the intuition behind human planning: near-term actions are concrete, while more distant plans remain provisional and are refined as new observations arrive. After the first action chunk is executed, the window advances, retaining the remaining predictions and appending a new chunk initialized with Gaussian noise. The retained predictions are then jointly refined with the new chunk, conditioned on the newly acquired camera observation. As chunks move through the window, their denoising steps are distributed across replanning cycles. Each steady-state cycle therefore requires only a fraction of the total steps to produce the next executable chunk. We train the model on matching noise profiles, with action tokens attending to partially denoised visual futures across the window. We evaluate Rolling-WAM on simulation benchmarks, including LIBERO [7] and RoboTwin [8], as well as three real-world humanoid manipulation tasks on the Unitree G1. Our experiments demonstrate that Rolling-WAM retains the benefits of visual imagination while significantly reducing computational overhead. It achieves a 98.1% average success rate on LIBERO and 93.3% on RoboTwin, remaining competitive with or exceeding state-of-the-art WAMs. Our primary contribution is a rolling formulation for joint video-action models that distributes denoising across successive replanning cycles. We demonstrate that this formulation achieves a 4.5×4.5× inference speedup over standard WAMs while maintaining competitive manipulation performance. Fig. 2: Rolling-WAM framework. (a) Video and action experts jointly predict flow velocities using masked joint attention and cross-attention to text and state context, conditioned on the current camera observation and per-chunk noise levels shared by video and actions. (b) Rolling inference denoises the window, executes the first action chunk, and shifts the retained future chunks while appending a new chunk initialized with Gaussian noise. (c) The same attention mask is used during training and inference. I Related Work I-A World Action Models for Robotic Manipulation World Action Models (WAMs) augment action generation with future visual prediction. Early approaches infer actions from language-conditioned video plans [9]. Recent methods integrate visual prediction and action generation using pretrained video models [1, 2, 10], often incorporating multimodal understanding [11, 3] or exploring native causal video-action pretraining [12]. Alternative designs learn predictive latent representations for action generation to avoid iterative future-video denoising at deployment [13, 14]. For joint WAMs, however, iterative video-action denoising remains computationally costly, motivating more efficient model designs and sampling strategies. I-B Efficient Inference for World Action Models Accelerating WAM inference typically involves reducing the computational burden of future prediction. Some methods remove future-video generation entirely at deployment [4] or extract future context in a single video-expert pass to cache for action denoising [15]. Other architectural optimizations include reducing model size, visual tokens, and video denoising steps [16], or accelerating sampling through modality-aware consistency distillation [17]. Other strategies focus on context reuse, either by recycling visual features after partial joint denoising [18] or by sharing asynchronously refreshed visual context across action updates [19]. Complementarily, real-time chunking overlaps inference with execution and uses action inpainting to align successive chunks [20]. Instead of removing or caching future predictions, Rolling-WAM distributes the joint video-action denoising process across replanning cycles, progressively refining near-term and future predictions within a sliding window. I-C Rolling Diffusion for Sequence Generation Rolling Diffusion [5] introduces a sliding denoising window with progressively higher noise levels toward the future. This concept has been extended with independent token noise levels for flexible sampling schedules [21] and bidirectional denoising with rollout-based distillation for autoregressive long-video generation [6]. These works motivate the rolling formulation of Rolling-WAM. In robotic manipulation, Streaming Diffusion Policy [22] and RNR-DP [23] maintain partially denoised action buffers to accelerate action generation. However, these methods focus exclusively on action-only denoising and evaluate on task-specific policies in relatively small-scale settings. We extend this rolling mechanism to couple world modeling with action generation, enabling action tokens to attend to partially denoised visual futures, and evaluate our approach across broader multitask settings. I Method This section presents Rolling-WAM, a joint video-action model for low-latency closed-loop robotic control. We first formulate the visual-action modeling problem and identify the computational bottleneck of standard joint denoising (Sec. I-A). We then introduce a rolling denoising mechanism that maintains a sliding window of predictions at staggered noise levels (Sec. I-B). Next, we describe the sampling and execution process that distributes computation across control cycles (Sec. I-C). Finally, we detail the model architecture and the training objective (Sec. I-D). I-A Problem Formulation We consider language-conditioned robotic manipulation from visual observations and proprioception. At time t, the policy receives an observation oto_t, robot state sts_t, and language instruction ℓ . It predicts an action sequence t:t+H−1a_t:t+H-1 over a horizon of H steps. World Action Models (WAMs) augment this action generation with future visual prediction [1, 2]. Let t+1:t+Hv_t+1:t+H denote the future video latents covering the corresponding physical interval at the video sampling rate. The joint visual-action modeling problem is formulated as learning the distribution: pθ(t:t+H−1,t+1:t+H∣ct),ct=(ot,st,ℓ).p_θ\! (a_t:t+H-1,v_t+1:t+H c_t ), c_t=(o_t,s_t, ). (1) This distribution can be modeled jointly or factorized with action generation conditioned on future visual prediction. Deploying WAMs in a closed loop requires multiple denoising steps within each replanning cycle before an action chunk is executed. This iterative process incurs a high computational cost. It increases inference latency and limits the responsiveness of the controller to new observations. As shown in Fig. 2, Rolling-WAM addresses this bottleneck by distributing the denoising computation across successive replanning cycles. It maintains a sliding window that jointly refines the next action chunk and its future continuation. This approach completes near-term predictions while leaving longer-term chunks partially denoised at staggered noise levels. As the window advances with execution, the retained predictions continue to evolve under new observations. This mechanism carries predictive context across chunk boundaries, allowing successive action chunks to be generated with a shared, evolving visual-action future. I-B Rolling Video-Action Denoising We partition the prediction window into W temporally aligned chunks 1:WX_1:W, where j=((j),(j))X_j=(v^(j),a^(j)) denotes the j-th video-action chunk. Each action chunk (j)=t+(j−1)K:t+jK−1a^(j)=a_t+(j-1)K:t+jK-1 contains K actions. Its corresponding video chunk (j)v^(j) covers the same physical interval at the designed video sampling rate. The total prediction horizon is therefore H=WKH=WK. Execution advances by one chunk at a time. During each replanning cycle, the window is processed jointly, but only the nearest chunk reaches completion. Let τ∈[0,1]τ∈[0,1] parameterize the denoising phase of a single replanning cycle, decreasing from 11 at the start to 00 at completion. Let σ(⋅)σ(·) denote a monotonically increasing base noise schedule, with σ(1)=1σ(1)=1 corresponding to Gaussian noise and σ(0)=0σ(0)=0 to clean data. We define two operational modes to establish and maintain the rolling window. Rolling mode The first chunk must become clean at the end of a replanning cycle. Each remaining chunk must be ready to occupy the preceding position in the window. We assign chunk j the noise level [5]: σjrolling(τ)=σ(j−1+τW),j=1,…,W. _j^rolling(τ)=σ\! ( j-1+τW ), j=1,…,W. (2) As τ decreases from 11 to 00, chunk j moves from noise level σ(j/W)σ(j/W) to σ((j−1)/W)σ((j-1)/W). The first chunk becomes clean, while later chunks remain at progressively higher noise levels. Crucially, σj+1rolling(0)=σjrolling(1),j=1,…,W−1. _j+1^rolling(0)= _j^rolling(1), j=1,…,W-1. (3) After the first chunk is removed, every retained chunk is already at the starting noise level for its new position. We append a new chunk initialized with Gaussian noise at σ=1σ=1, restoring the configuration for the next replanning cycle. Initialization mode At the beginning of an episode, partially denoised predictions are unavailable. We initialize the entire window from Gaussian noise and use the schedule: σjinit(τ)=σ(min1,τ+j−1W). _j^init(τ)=σ\! ( \1,\;τ+ j-1W \ ). (4) All chunks start at a noise level σ=1σ=1. Nearer chunks begin denoising earlier, while more distant chunks remain at pure noise until their refinement begins. At completion, σjinit(0)=σjrolling(0) _j^init(0)= _j^rolling(0). The first chunk is executable, and the remaining predictions enter rolling mode after the window shifts. Together, these two modes allow each chunk to begin as a distant prediction and receive further refinement as it approaches execution. Video and actions within each chunk share the same denoising steps. I-C Sampling and Execution We allocate N total denoising steps per chunk. We choose N as a multiple of W to ensure an even distribution across rolling cycles. Initialization uses Δτ=−1/N τ=-1/N at each denoising step, requiring N steps to reach rolling mode. Subsequent replanning cycles require only N/WN/W steps to produce the next executable action chunk, with Δτ=−W/N τ=-W/N at each step. At each denoising step, the model jointly updates the video and action predictions throughout the window. Let ~j X_j denote the current denoising state of chunk j. Given the predicted flow velocity fθ,jf_θ,j, the Euler update is: ~j←~j+[j(τ+Δτ)−j(τ)]fθ,j(~1:W,;ct). X_j← X_j+ [ σ_j(τ+ τ)- σ_j(τ) ]f_θ,j\! ( X_1:W, σ;c_t ). (5) Once the first chunk is clean, the robot executes its K actions. The window then slides over. It removes the executed chunk, retains the remaining predictions, and appends a new chunk initialized with Gaussian noise (Fig. 2b). The robot acquires the latest camera observation and state to condition the next replanning cycle. Each chunk traverses all N denoising steps before execution, but its computation is distributed across multiple replanning cycles. This mechanism reduces the number of denoising steps needed to produce the next executable chunk, while preserving an evolving visual-action prediction across chunk boundaries. Consequently, larger windows can lower steady-state inference latency, assuming fixed N and sufficient GPU parallelism to process the extended prediction window. This approach is orthogonal to other efficiency techniques [15, 16] for WAMs and can be combined to further reduce inference latency. I-D Model Architecture and Training As shown in Fig. 2(a), Rolling-WAM pairs a pretrained video Diffusion Transformer [24, 25] with a lightweight action Transformer in a Mixture-of-Transformers (MoT) architecture [4, 26]. The video expert processes VAE latents of the current observation and noisy future video. The action expert processes projected noisy actions. Masked joint attention couples the experts at each layer, allowing action generation to draw on pretrained visual dynamics. Language and proprioceptive embeddings condition both experts through cross-attention. To support rolling denoising, video and action tokens are modulated by their respective chunk-wise noise levels. This allows predictions at different stages of refinement to be processed jointly. The attention mask (Fig. 2c) allows every action chunk to attend to visual features across the entire prediction window. Direct action-to-action attention is restricted to the same chunk, and video tokens do not attend to action tokens. Each action chunk therefore draws on a shared, evolving visual prediction that extends beyond its own execution interval. Training objective. We train the joint denoiser with flow matching [27]. For each demonstration window, we sample τ∼(0,1)τ (0,1) and select rolling or initialization mode with probabilities β and 1−β1-β, respectively. Let j σ_j denote the video and action noise levels of chunk j under selected schedule. Let =(1,…,W) σ=( σ_1,…, σ_W) denote the window noise profile. We construct the noisy targets: ~j=(1−j)j+jϵj,ϵj∼(,), X_j=(1- σ_j)X_j+ σ_j ε_j, ε_j (0,I), (6) using independent Gaussian samples for video and actions. The network predicts conditional velocities for all chunks, with target ϵj−j ε_j-X_j. For modality q∈v,aq∈\v,a\, the per-chunk flow-matching loss is: ℓjq=‖[fθq(~1:W,;ct)]j−(ϵjq−jq)‖22, _j^q= \| [f_θ^q\! ( X_1:W, σ;c_t ) ]_j- ( ε_j^q-X_j^q ) \|_2^2, (7) where [⋅]j[·]_j selects the prediction for chunk j, jv=(j)X_j^v=v^(j), and ja=(j)X_j^a=a^(j). The joint training objective is: ℒ=[1W∑j=1Wbjw(j)(λvℓjv+λaℓja)],L=E\! [ 1W _j=1^Wb_j\,w( σ_j) ( _v _j^v+ _a _j^a ) ], (8) where w weights the noise levels and λv,λa _v, _a balance the two modalities. The activity mask bjb_j excludes initialization chunks held at pure noise. Training on both modes equips the same model to initialize the prediction window and continue its refinement during closed-loop execution. IV Experiments Our experiments verify whether rolling denoising can accelerate joint video-action prediction while preserving manipulation performance. We compare task success with a broad set of vision-language-action models and world action models on LIBERO, RoboTwin 2.0, and real-world Unitree G1 tasks. Controlled latency measurements quantify the computational savings of rolling denoising. Ablations examine the roles of window size, training noise schedules, and cross-chunk action attention. IV-A Experimental Setup We evaluate Rolling-WAM on the following settings. We report the task success rate (%) as the primary performance metric and measure replanning latency to assess efficiency. IV-A1 LIBERO We evaluate the Spatial, Object, Goal, and Long suites of LIBERO [7], each containing 10 tasks and 500 demonstrations. Observations include external and wrist RGB views. We train one policy across all four suites for 10 epochs with an effective batch size of 128, and evaluate each task with 50 rollouts. IV-A2 RoboTwin 2.0 We evaluate 50 bimanual manipulation tasks in the Clean and Randomized settings of RoboTwin 2.0 [8]. Observations include RGB images from the head and two wrist cameras. We train one policy across 50 tasks using 2,500 clean demonstrations and 25,000 demonstrations from randomized scenes, for 5 epochs with an effective batch size of 1024. Each task is evaluated with 100 rollouts per setting. IV-A3 Real-world humanoid manipulation As shown in Fig. 6, we evaluate three tasks on a Unitree G1 humanoid: Doll Placement (placing a plush dog in a box), Plate Stacking (stacking three plates), and Bead Pouring (transferring beads from a bottle into a glass). The policy receives a 320×224320× 224 egocentric RGB image and a 43-dimensional robot state. It predicts 78-dimensional actions comprising a 64-dimensional SONIC motion latent [28] and 7-dimensional commands for each hand. We train one policy across all three tasks using 50 demonstrations per task, recorded at 10 Hz. Training runs for 7,500 steps with an effective batch size of 192. We execute actions at 10 Hz and evaluate 20 trials per task. IV-B Implementation Details We initialize the video expert from Wan2.2-TI2V-5B [24] and reuse its pretrained text encoder and VAE. The action expert has 30 layers and hidden dimension da=1024d_a=1024 (approximately 1B parameters). Its backbone is initialized by interpolating the video expert’s weights [4]. We jointly train both experts and the proprioceptive encoder, while freezing the VAE and text encoder. All training uses benchmark demonstrations without additional embodied pretraining. We use AdamW with learning rate 10−410^-4, weight decay 10−210^-2, 5% linear warmup followed by cosine decay, and BF16 mixed precision. Video and action losses are averaged separately, with λv=λa=1 _v= _a=1. Initialization and rolling modes are sampled with probabilities 0.2 and 0.8. Both modalities use the same shifted noise schedule [29], σ(τ)=ρτ/[1+(ρ−1)τ]σ(τ)=ρτ/[1+(ρ-1)τ] with ρ=5ρ=5, during training and inference. Unless otherwise specified, we use N=10N=10 denoising steps per chunk, a window of W=5W=5 chunks, and K=16K=16 actions per chunk. Each steady-state replanning cycle therefore uses N/W=2N/W=2 denoising steps. The classifier-free guidance scale is 1. TABLE I: Success rates (%) on LIBERO. P.T. indicates embodied pretraining. The highest and second-highest average scores are bold and underlined, respectively. Rolling-WAM achieves competitive performance with faster inference. Method P.T. Spatial Object Goal Long Average π0 _0 [30] ✓ 96.8 98.8 95.8 85.2 94.1 π0.5 _0.5 [31] ✓ 98.8 98.2 98.0 92.4 96.9 Motus [11] ✓ 96.8 99.8 96.6 97.6 97.7 LingBot-VA [1] ✓ 98.5 99.6 97.2 98.5 98.5 Fast-WAM [4] ✗ 98.2 100.0 97.0 95.2 97.6 Joint-WAM [4] ✗ 99.6 99.4 98.2 96.8 98.5 Rolling-WAM (Ours) ✗ 98.2 98.0 98.2 97.8 98.1 TABLE I: Success rates (%) on RoboTwin 2.0. P.T. indicates embodied pretraining. The highest and second-highest average scores are bold and underlined, respectively. Rolling-WAM achieves the highest success rate in both settings. Method P.T. Clean Rand. Average π0 _0 [30] ✓ 65.9 58.4 62.2 π0.5 _0.5 [31] ✓ 82.7 76.8 79.8 Motus [11] ✓ 88.7 87.0 87.8 LingBot-VA [1] ✓ 92.9 91.5 92.2 Fast-WAM [4] ✗ 91.9 91.8 91.8 Joint-WAM [4] ✗ 90.8 90.3 90.6 Rolling-WAM (Ours) ✗ 93.5 93.0 93.3 Baselines. We compare with both VLA and WAM policies, including π0 _0 [30], π0.5 _0.5 [31], GR00T N1.7 [32], Motus [11], LingBot-VA [1], Fast-WAM [4], and Joint-WAM [4]. For controlled WAM comparisons, we reproduce Fast-WAM and Joint-WAM using matched training and evaluation settings wherever applicable. Joint-WAM jointly denoises future video and actions from Gaussian noise at each replanning cycle, while Fast-WAM generates actions without future-video denoising at deployment. IV-C Performance on Simulation Benchmarks Fig. 3: Imagined and observed rollouts. Rolling-WAM’s imagined video (top) and ground-truth observations (bottom) during policy execution on Put Object Cabinet. Columns show successive moments. We first evaluate whether rolling denoising can deliver competitive manipulation performance with fewer denoising steps per replanning cycle. As shown in Table I, Rolling-WAM achieves an average success rate of 98.1% on LIBERO, within 0.4 percentage points of the joint leaders, LingBot-VA and Joint-WAM (both 98.5%), and above Motus (97.7%) and Fast-WAM (97.6%). Success ranges from 97.8% to 98.2% across the four suites, showing consistently strong performance with rolling inference. RoboTwin 2.0 further tests bimanual manipulation under scene randomization. As shown in Table I, Rolling-WAM achieves success rates of 93.5% and 93.0% in clean and randomized settings, respectively. Its 93.3% average leads LingBot-VA (92.2%), Fast-WAM (91.8%), and Joint-WAM (90.6%). The high success rate in both settings shows that rolling inference remains effective under the evaluated scene variations, without additional embodied pretraining. See Appendix Table V for per-task results. Rolling-WAM achieves these results with only two denoising steps per replanning cycle, suggesting that co-refining current and future predictions can potentially combine faster replanning with improved manipulation performance. We also inspect whether the retained visual predictions remain aligned with observations as execution proceeds. Figure 3 compares imagined and observed rollouts on Put Object Cabinet in the RoboTwin 2.0 benchmark. Imagined robot motion and task progression closely follow the observations in this example, providing a qualitative view of the visual context maintained during rolling execution. IV-D Inference Efficiency To formally investigate the inference efficiency gains of rolling denoising, we quantitatively compare Rolling-WAM with the baselines and examine how replanning latency varies with the prediction horizon. We measure steady-state replanning latency on a single NVIDIA A100 GPU under the RoboTwin 2.0 setting at an image resolution of 384×320384× 320. All three methods execute 16 actions per replanning cycle. Timing includes visual encoding and denoising, with CUDA synchronization, and excludes warm-up and initialization. Fig. 4: Replanning latency comparison. Average latency over multiple steady-state replanning cycles, excluding initialization. Lower is better. Default configuration. As shown in Fig. 4, with N=10N=10 and W=5W=5, Rolling-WAM uses two denoising steps per update and takes 215 ms, compared with 978 ms for Joint-WAM and 548 ms for Fast-WAM. These correspond to approximately 4.5×4.5× and 2.5×2.5× speedups, respectively. The VLA baselines π0.5 _0.5 and GR00T N1.7 take 296 ms and 285 ms per update, respectively. Rolling-WAM thus retains future-video generation while taking less time per update than either VLA baseline in this comparison. In the meantime, Rolling-WAM retains an 80-action prediction window, while the baselines predict 16 actions. The reduction in latency is therefore achieved while refining a longer future at each update. These timings use no torch.compile, TensorRT, or custom CUDA kernels. Window-size scaling. To examine how distributing denoising over more cycles affects latency, we vary the window size W from 1 to 8. To match execution horizon, Rolling-WAM uses H=16WH=16W, while the baselines use H=16H=16. At each W, all methods use N=W⌊16/W⌋N=W 16/W total denoising steps. Rolling inference uses N/W=⌊16/W⌋N/W= 16/W steps per replanning cycle. We set the denoising step as 16 to keep rounding effects small across the tested window sizes. As shown in Fig. 5, at W=5W=5, Rolling-WAM requires only 322 ms, compared with 832 ms for Fast-WAM and 1529 ms for Joint-WAM, yielding 2.58×2.58× and 4.75×4.75× speedups, respectively. The latency curve flattens for W=6W=6–8, where the number of rolling denoising steps stays at two per replanning cycle. We use W=5W=5 by default to balance latency and task performance (Sec. IV-F). Fig. 5: Replanning latency across window sizes. At each window size W, Rolling-WAM and the baselines use prediction horizons H=16WH=16W and H=16H=16, respectively, with the total denoising budget N=W⌊16/W⌋N=W 16/W . Points show mean replanning latency (ms), excluding initialization. Fig. 6: Real-world humanoid manipulation tasks. Doll Placement, Plate Stacking, and Bead Pouring (top to bottom), illustrated with qualitative examples from Rolling-WAM rollouts. Each row shows task progression from left to right. Insets show egocentric observations. IV-E Real-World Humanoid Evaluation To assess the practical benefits of rolling denoising in physical manipulation, we evaluate Rolling-WAM on real-world humanoid tasks. A trial succeeds when the robot completes the task without human intervention: placing the doll in the box, stacking all three plates, or transferring beads into the glass. Figure 6 shows representative rollouts. As shown in Table I, Rolling-WAM achieves the highest average success rate of 85.0% among the five policies, followed by Joint-WAM at 78.3% and Fast-WAM at 75.0%. It matches the strongest baselines on Doll Placement (85%) and Plate Stacking (100%), while achieving the highest success rate on Bead Pouring (70%). These results are obtained with a single policy trained across all three tasks. Qualitatively, we observe more continuous execution across action-chunk boundaries with Rolling-WAM. Pauses are more apparent with Joint-WAM and sometimes interrupt task progress. These observations illustrate the practical importance of replanning latency during physical manipulation. TABLE I: Humanoid manipulation on Unitree G1. Each task is evaluated over 20 rollouts per method. Task cells report success rates (%). The highest and second-highest average scores are bold and underlined, respectively. Method Doll Placement Plate Stacking Bead Pouring Avg. (%) π0.5 _0.5 [31] 55.0 70.0 60.0 61.7 GR00T N1.7 [32] 75.0 65.0 60.0 66.7 Fast-WAM [4] 85.0 80.0 60.0 75.0 Joint-WAM [4] 70.0 100.0 65.0 78.3 Rolling-WAM (Ours) 85.0 100.0 70.0 85.0 IV-F Ablation Studies Fig. 7: Window size and task success. Average success rates on the six selected RoboTwin tasks under the Clean setting. The prediction horizon is H=16WH=16W. TABLE IV: Ablations on noise scheduling and action attention. Mean success rates (%) on six RoboTwin tasks and four LIBERO suites. A2A denotes bidirectional action attention across chunks; bold indicates the best per column. Variant Selected RoboTwin LIBERO Full Rolling (w/o A2A) 78.2 98.1 w/ A2A 76.3 97.9 Constant (p=0.2p=0.2) 74.7 97.9 Constant (p=0.5p=0.5) 73.0 97.1 Random (p=0.2p=0.2) 75.3 97.0 Random (p=0.5p=0.5) 78.5 97.3 To support our design choices, we examine the prediction window, training noise schedules, and action attention. Table IV reports results on LIBERO and six representative RoboTwin Clean tasks: Lift Pot, Beat Block Hammer, Place Dual Shoes, Stack Bowls Two, Blocks Ranking Size, and Stack Blocks Three. Window size. Increasing W provides a longer predicted future, but its value depends on whether the additional context helps the next action. To test this, we vary W=1,…,8W=1,…,8 on the six selected RoboTwin tasks, with H=16WH=16W and N=W⌊16/W⌋N=W 16/W , following the sampling rule in Sec. IV-D. As shown in Fig. 7, success peaks at 78.2% for W=5W=5, versus 77.3% for W=3W=3, and 69.5% for W=8W=8. Larger windows do not consistently improve performance. We conjecture that more distant visual predictions are less constrained by the current observation and may offer limited guidance for the imminent action chunk. The best result in this sweep comes from a moderate window, which motivates our choice of W=5W=5 as the default. Training noise mixture. Our default training matches the staggered noise profiles used at inference. We test whether exposing the model to a broader mixture improves policy performance. We keep initialization training unchanged and replace a fraction p of rolling-mode samples with alternative noise profiles. Constant uses one shared noise level across chunks; Random samples a separate level for each chunk. The results are shown in Table IV. The full rolling schedule gives 78.2% success on selected RoboTwin tasks and 98.1% on LIBERO. Random with p=0.5p=0.5 raises RoboTwin success to 78.5%, but lowers LIBERO to 97.3%. No alternative improves both benchmarks. The full rolling schedule achieves the best LIBERO score and is within 0.3 percentage points of the best RoboTwin variant. We therefore retain the rolling schedule for all non-initialization samples. Cross-chunk action attention. Action chunks already receive context from the entire visual window. We test whether adding direct attention between action chunks further improves performance. As shown in Table IV, allowing this attention gives 76.3% success on the selected RoboTwin tasks and 97.9% on LIBERO, compared with 78.2% and 98.1% with within-chunk action attention. These results favor sharing information through the visual window while restricting direct action attention to each chunk. V Conclusion In this paper, we presented Rolling-WAM, an efficient World Action Model that jointly refines near-term and future video-action predictions within a sliding window. By distributing denoising across replanning cycles, it reduces the sequential steps needed to produce the next executable action chunk, while using partially denoised visual futures to guide action generation. Experiments show that Rolling-WAM achieves competitive performance on LIBERO and outperforms the compared methods on RoboTwin 2.0 and real-world humanoid manipulation tasks, while delivering substantial speedups in inference over standard joint WAMs. Combining rolling denoising with complementary model and system optimizations offers further opportunities for low-latency robot control. Limitations and future work. Design choices such as window and chunk sizes warrant further exploration across task settings with different dynamics and control frequencies. Despite observation feedback, retained predictions may lag behind rapid scene changes and misguide action generation, particularly with long windows. Adaptive window management and asynchronous execution offer directions for improving closed-loop responsiveness. Acknowledgments This work was partially supported by the National Science Foundation through NSF CPS #2434460. The USC Physical Superintelligence Lab acknowledges generous support from Toyota Research Institute, Dolby, Google DeepMind, Capital One, Nvidia, Bosch, NSF, and Qualcomm. Yue Wang is also supported by a Powell Research Award. References [1] L. Li, Q. Zhang, Y. Luo, S. Yang, R. Wang, F. Han, M. Yu, Z. Gao, N. Xue, X. Zhu, Y. Shen, and Y. Xu (2026) Causal World Modeling for Robot Control. In Proceedings of the Robotics: Science and Systems Conference (RSS), Cited by: §I, §I-A, §I-A, §IV-B, TABLE I, TABLE I. [2] S. Ye, Y. Ge, K. Zheng, S. Gao, S. Yu, G. Kurian, S. Indupuru, Y. L. Tan, C. Zhu, J. Xiang, A. N. Malik, K. Lee, W. Liang, N. R. Arachchige, J. Gu, Y. Xu, G. Wang, F. Hu, A. Narayan, J. Bjorck, J. Wang, G. Kim, D. Niu, R. Zheng, Y. Xie, J. Wu, Q. Wang, D. Xu, Y. Du, R. Julian, Y. Chebotar, S. Reed, J. Kautz, Y. Zhu, L. Fan, and J. Jang (2026) World Action Models are Zero-shot Policies. In Proceedings of the Conference on Robot Learning (CoRL), Cited by: §I, §I, §I-A, §I-A. [3] NVIDIA (2026) Cosmos 3: Omnimodal World Models for Physical AI. arXiv preprint arXiv:2606.02800. Cited by: §I, §I-A. [4] T. Yuan, Z. Dong, Y. Liu, and H. Zhao (2026) Fast-WAM: Do World Action Models Need Test-time Future Imagination?. arXiv preprint arXiv:2603.16666. Cited by: §I, §I, §I-B, §I-D, §IV-B, §IV-B, TABLE I, TABLE I, TABLE I, TABLE I, TABLE I, TABLE I. [5] D. Ruhe, J. Heek, T. Salimans, and E. Hoogeboom (2024) Rolling Diffusion Models. In Proceedings of the International Conference on Machine Learning (ICML), Cited by: §I, §I-C, §I-B. [6] K. Liu, W. Hu, J. Xu, Y. Shan, and S. Lu (2026) Rolling Forcing: Autoregressive Long Video Diffusion in Real Time. In Proceedings of the International Conference on Learning Representations (ICLR), Cited by: §I, §I-C. [7] B. Liu, Y. Zhu, C. Gao, Y. Feng, Q. Liu, Y. Zhu, and P. Stone (2023) LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learning. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), Cited by: §I, §IV-A1. [8] T. Chen, Z. Chen, B. Chen, Z. Cai, Y. Liu, Z. Li, Q. Liang, X. Lin, Y. Ge, Z. Gu, W. Deng, Y. Guo, T. Nian, X. Xie, Q. Chen, K. Su, T. Xu, G. Liu, M. Hu, H. Gao, K. Wang, Z. Liang, Y. Qin, X. Yang, P. Luo, and Y. Mu (2026) RoboTwin 2.0: A Scalable Data Generator and Benchmark with Strong Domain Randomization for Robust Bimanual Robotic Manipulation. In Proceedings of the International Conference on Machine Learning (ICML), Cited by: §I, §IV-A2. [9] Y. Du, S. Yang, B. Dai, H. Dai, O. Nachum, J. Tenenbaum, D. Schuurmans, and P. Abbeel (2023) Learning Universal Policies via Text-Guided Video Generation. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), Cited by: §I-A. [10] M. J. Kim, Y. Gao, T. Lin, Y. Lin, Y. Ge, G. Lam, P. Liang, S. Song, M. Liu, C. Finn, and J. Gu (2026) Cosmos Policy: Fine-Tuning Video Models for Visuomotor Control and Planning. In Proceedings of the International Conference on Learning Representations (ICLR), Cited by: §I-A. [11] H. Bi, H. Tan, S. Xie, Z. Wang, S. Huang, H. Liu, R. Zhao, Y. Feng, C. Xiang, Y. Rong, H. Zhao, H. Liu, Z. Su, L. Ma, H. Su, and J. Zhu (2026) Motus: A Unified Latent Action World Model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Cited by: §I-A, §IV-B, TABLE I, TABLE I. [12] Q. Zhang, L. Li, L. Zhang, S. Yang, Y. Luo, S. Li, R. Wang, J. Wang, J. Shao, G. Xu, J. Zhou, Y. Shen, Y. Jin, F. Xu, S. Ma, J. Liao, G. Lu, Z. Shi, Y. Wen, Y. Zhao, W. Tang, X. Wang, C. Li, J. Zhu, K. L. Cheng, N. Xue, X. Zhu, Y. Shen, and Y. Xu (2026) Native Video-Action Pretraining for Generalizable Robot Control. arXiv preprint arXiv:2607.08639. Cited by: §I-A. [13] H. Ma, J. Cai, X. Xu, H. Li, Y. Yang, Y. Tian, J. Cao, H. Zhu, Z. Qiu, Zhaxizhuoma, Y. Yang, J. Peng, X. Wei, Y. Zhu, J. Jiang, X. Gao, H. Wang, F. Yuan, K. Li, X. Zhu, T. Wang, Y. Ding, J. Pang, J. Zeng, J. Zhang, B. Zhou, Y. Mu, C. Shen, and W. Zhang (2026) InternVLA-A1.5: Unifying Understanding, Latent Foresight, and Action for Compositional Generalization. arXiv preprint arXiv:2607.04988. Cited by: §I-A. [14] Z. Li, Z. Zhang, Y. Wei, W. Zhang, X. Yuan, P. Zhi, G. Li, X. Guo, F. Gao, J. Yang, and S. Zhang (2026) ω-0: A Latent Predictive World Action Model for Concurrent Humanoid Loco-Manipulation. arXiv preprint arXiv:2608.06375. Cited by: §I-A. [15] W. Zhao, H. Jiang, X. Shi, L. Liu, F. Huang, Z. Su, W. Sui, and X. Wang (2026) Faster-WAM: Efficient Inference-Time Future Conditioning for Robust World Action Models. arXiv preprint arXiv:2608.04404. Cited by: §I-B, §I-C. [16] J. Li, T. Guo, Y. Ye, R. Zhang, X. Chi, Q. Sun, Y. Li, Y. Lou, Y. Huang, Z. Lu, M. Guo, and S. Zhang (2026) Efficient-WAM: A 1B-Parameter World-Action Model with Low-Cost Future Imagination. arXiv preprint arXiv:2606.10040. Cited by: §I-B, §I-C. [17] A. Akbari, C. Zhang, A. Akbari, L. Zhao, Y. Chen, W. Chen, X. Zhang, G. Yuan, and Y. Wang (2026) Flash-WAM: Modality-Aware Distillation for World Action Models. arXiv preprint arXiv:2606.05254. Cited by: §I-B. [18] C. Xiang, F. Bao, H. Liu, H. Tan, H. Bi, J. Li, J. Liu, J. Pang, K. Jing, L. Liu, M. Cai, R. Cui, R. Zhao, R. Wang, S. Huang, Y. Feng, Y. Rong, Z. Wang, and J. Zhu (2026) Motubrain: An Advanced World Action Model for Robot Control. arXiv preprint arXiv:2604.27792. Cited by: §I-B. [19] J. Cai, L. Ling, S. Chu, Z. Liu, J. Kang, Z. Liang, W. Xu, Y. Mao, W. Zhang, X. Yang, R. Ying, R. Zheng, and Y. Mu (2026) AHA-WAM: Asynchronous Horizon-Adaptive World-Action Modeling with Observation-Guided Context Routing. arXiv preprint arXiv:2606.09811. Cited by: §I-B. [20] K. Black, M. Y. Galliker, and S. Levine (2025) Real-Time Execution of Action Chunking Flow Policies. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), Cited by: §I-B. [21] B. Chen, D. Martí Monsó, Y. Du, M. Simchowitz, R. Tedrake, and V. Sitzmann (2024) Diffusion Forcing: Next-token Prediction Meets Full-Sequence Diffusion. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), Cited by: §I-C. [22] S. H. Høeg, Y. Du, and O. Egeland (2025) Streaming Diffusion Policy: Fast Policy Synthesis with Variable Noise Diffusion Models. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), Cited by: §I-C. [23] Z. Chen, X. Yuan, T. Mu, and H. Su (2025) Responsive Noise-Relaying Diffusion Policy: Responsive and Efficient Visuomotor Control. Transactions on Machine Learning Research. Cited by: §I-C. [24] Team Wan et al. (2025) Wan: Open and Advanced Large-Scale Video Generative Models. arXiv preprint arXiv:2503.20314. Cited by: §I-D, §IV-B. [25] W. Peebles and S. Xie (2023) Scalable Diffusion Models with Transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), Cited by: §I-D. [26] W. Liang, L. Yu, L. Luo, S. Iyer, N. Dong, C. Zhou, G. Ghosh, M. Lewis, W. Yih, L. Zettlemoyer, and X. V. Lin (2025) Mixture-of-Transformers: A Sparse and Scalable Architecture for Multi-Modal Foundation Models. Transactions on Machine Learning Research. Cited by: §I-D. [27] Y. Lipman, R. T. Q. Chen, H. Ben-Hamu, M. Nickel, and M. Le (2023) Flow Matching for Generative Modeling. In Proceedings of the International Conference on Learning Representations (ICLR), Cited by: §I-D. [28] Z. Luo, Y. Yuan, T. Wang, C. Li, F. Castañeda, S. Chen, Z. Cao, J. Li, D. Minor, Q. Ben, J. Park, D. Sami, Z. Wang, X. Da, R. Ding, C. Hogg, L. Song, E. Lim, E. Jeong, T. He, H. Xue, W. Xiao, S. Yuen, J. Kautz, Y. Chang, U. Iqbal, L. “. Fan, and Y. Zhu (2026) SONIC: Supersizing motion tracking for natural humanoid whole-body control. Science Robotics 11 (117). Cited by: §IV-A3. [29] P. Esser, S. Kulal, A. Blattmann, R. Entezari, J. Müller, H. Saini, Y. Levi, D. Lorenz, A. Sauer, F. Boesel, D. Podell, T. Dockhorn, Z. English, and R. Rombach (2024) Scaling Rectified Flow Transformers for High-Resolution Image Synthesis. In Proceedings of the International Conference on Machine Learning (ICML), Cited by: §IV-B. [30] K. Black, N. Brown, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusai, L. Groom, K. Hausman, B. Ichter, S. Jakubczak, T. Jones, L. Ke, S. Levine, A. Li-Bell, M. Mothukuri, S. Nair, K. Pertsch, L. X. Shi, J. Tanner, Q. Vuong, A. Walling, H. Wang, and U. Zhilinsky (2025) π0 _0: A Vision-Language-Action Flow Model for General Robot Control. In Proceedings of the Robotics: Science and Systems Conference (RSS), Cited by: §IV-B, TABLE I, TABLE I. [31] K. Black, N. Brown, J. Darpinian, K. Dhabalia, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusai, M. Y. Galliker, D. Ghosh, L. Groom, K. Hausman, B. Ichter, S. Jakubczak, T. Jones, L. Ke, D. LeBlanc, S. Levine, A. Li-Bell, M. Mothukuri, S. Nair, K. Pertsch, A. Z. Ren, L. X. Shi, L. Smith, J. T. Springenberg, K. Stachowicz, J. Tanner, Q. Vuong, H. Walke, A. Walling, H. Wang, L. Yu, and U. Zhilinsky (2025) π0.5 _0.5: A Vision-Language-Action Model with Open-World Generalization. In Proceedings of the Conference on Robot Learning (CoRL), Cited by: §IV-B, TABLE I, TABLE I, TABLE I. [32] J. Bjorck, F. Castañeda, N. Cherniadev, X. Da, R. Ding, L. ”. Fan, Y. Fang, D. Fox, F. Hu, S. Huang, J. Jang, Z. Jiang, J. Kautz, K. Kundalia, L. Lao, Z. Li, Z. Lin, K. Lin, G. Liu, E. Llontop, L. Magne, A. Mandlekar, A. Narayan, S. Nasiriany, S. Reed, Y. L. Tan, G. Wang, Z. Wang, J. Wang, Q. Wang, J. Xiang, Y. Xie, Y. Xu, Z. Xu, S. Ye, Z. Yu, A. Zhang, H. Zhang, Y. Zhao, R. Zheng, and Y. Zhu (2025) GR00T N1: An Open Foundation Model for Generalist Humanoid Robots. arXiv preprint arXiv:2503.14734. Cited by: §IV-B, TABLE I. [RoboTwin Detailed Results] Table V presents per-task success rates for Rolling-WAM and the baselines on all 50 RoboTwin 2.0 tasks under clean and randomized evaluation settings. TABLE V: Per-task success rates (%) on RoboTwin 2.0 under clean and randomized settings. Bold denotes the best score in each setting. Task Rolling-WAM Fast-WAM Joint-WAM LingBot-VA π0.5 _0.5 Motus Clean Rand. Clean Rand. Clean Rand. Clean Rand. Clean Rand. Clean Rand. Adjust Bottle 99 97 100 100 98 99 90 94 100 99 89 93 Beat Block Hammer 91 91 99 97 100 98 96 98 96 93 95 88 Blocks Ranking RGB 97 95 100 100 100 100 99 98 92 85 99 97 Blocks Ranking Size 83 87 94 98 83 91 94 96 49 26 75 63 Click Alarmclock 100 99 100 100 100 100 99 100 98 89 100 100 Click Bell 100 100 100 100 100 98 100 100 99 66 100 100 Dump Bin Bigbin 95 96 97 96 95 95 89 96 92 97 95 91 Grab Roller 100 100 100 100 100 100 100 100 100 100 100 100 Handover Block 94 93 95 81 93 91 99 78 66 57 86 73 Handover Mic 95 100 99 100 100 100 94 96 98 97 78 63 Hanging Mug 70 58 58 62 71 56 40 28 18 17 38 38 Lift Pot 100 100 100 100 100 100 100 99 96 85 96 99 Move Can Pot 98 97 90 88 97 99 94 97 51 55 34 74 Move Pillbottle Pad 100 99 100 99 99 100 99 99 84 61 93 96 Move Playingcard Away 100 97 100 100 100 100 100 99 96 84 100 96 Move Stapler Pad 83 84 77 64 85 81 91 79 56 42 83 85 Open Laptop 97 98 98 100 89 92 92 94 90 96 95 91 Open Microwave 100 91 62 45 3 14 82 86 34 77 95 91 Pick Diverse Bottles 85 87 80 85 86 87 89 82 81 71 90 91 Pick Dual Bottles 98 99 100 96 98 99 100 99 93 63 96 90 Place A2B Left 97 92 95 93 96 96 97 93 87 82 88 79 Place A2B Right 98 97 93 99 95 95 97 95 87 84 91 87 Place Bread Basket 97 93 91 93 89 94 97 95 77 64 91 94 Place Bread Skillet 93 98 90 93 90 93 95 90 85 66 86 83 Place Burger Fries 97 98 96 99 100 100 97 95 94 87 98 98 Place Can Basket 76 70 71 69 50 23 81 84 62 62 81 76 Place Cans Plasticbox 100 99 99 96 98 98 100 99 94 84 98 94 Place Container Plate 97 98 96 100 99 98 99 97 99 95 98 99 Place Dual Shoes 90 91 94 88 93 89 94 89 75 75 93 87 Place Empty Cup 100 100 100 100 100 100 100 100 100 99 99 98 Place Fan 93 93 96 96 99 96 99 93 87 85 91 87 Place Mouse Pad 97 98 83 89 96 91 93 96 60 39 66 68 Place Object Basket 95 86 89 88 86 81 91 88 80 76 81 87 Place Object Scale 91 98 90 97 96 99 96 95 86 80 88 85 Place Object Stand 90 93 90 94 92 98 99 96 91 85 98 97 Place Phone Stand 100 97 97 99 100 100 97 97 81 81 87 86 Place Shoe 95 98 96 99 95 97 98 98 92 93 99 97 Press Stapler 83 79 90 97 52 50 85 82 87 83 93 98 Put Bottles Dustbin 96 97 95 90 93 95 87 91 84 79 81 79 Put Object Cabinet 86 91 94 89 95 90 85 87 80 79 88 71 Rotate QRcode 88 78 93 89 91 92 96 91 89 87 89 73 Scan Object 95 90 89 92 92 92 96 91 72 65 67 66 Shake Bottle 100 99 100 100 100 100 100 97 99 97 100 97 Shake Bottle Horizontally 100 100 100 100 100 100 100 99 99 99 100 98 Stack Blocks Three 99 98 95 97 98 97 99 98 91 76 91 95 Stack Blocks Two 100 100 100 100 100 100 100 98 97 100 100 98 Stack Bowls Three 83 82 80 81 84 86 86 83 77 71 79 87 Stack Bowls Two 95 97 92 98 97 95 94 98 95 96 98 98 Stamp Seal 96 98 90 94 96 99 96 97 79 55 93 92 Turn Switch 65 73 61 59 73 72 44 45 62 54 84 78 Average 93.54 92.98 91.88 91.78 90.84 90.32 92.90 91.50 82.74 76.76 88.66 87.02