acceptodds
Under review as a conference paper at ICLR 2027

DiSpecWAM: Distillation and Speculative Inference for World Action Models

Abstract

Current World Action Models (WAMs) face a fundamental tension between world-modeling capability and real-time interaction efficiency. Future video prediction helps model environment dynamics and reason about the consequences of actions, but multi-step video–action generation incurs substantial inference overhead. Existing methods either remove video prediction at the cost of essential world-modeling capabilities, or aggressively compress the generation process, making it difficult to preserve the performance of the full model. To address capability preservation and error recovery in few-step generation, we propose DiSpecWAM. First, F²Distill converts a full-capacity WAM into a complementary hierarchy comprising a low-budget Draft Model and a higher-fidelity Target Model. By coupling teacher-flow preservation with Student-rollout distribution calibration, both retain reliable visual foresight and precise action generation despite extreme denoising reduction. Then, R²Spec organizes them into repairable and reliable speculative inference, where the Target Model selectively verifies, repairs, and falls back on unreliable Draft Model outputs. This design couples high-fidelity compression during training with adaptive recovery during inference. On RoboTwin 2.0, DiSpecWAM achieves 92.65% average success at 205 ms per action chunk, delivering a 16.61× speedup over the full WAM.

Primary area:
applications to robotics, autonomy, planning
Keywords:
WAM, Inference-speedup

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_cchLOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper speeds up inference for world action models, which jointly generate future video and robot actions, while keeping explicit video prediction at test time. F2 Distill compresses a full-budget LingBot-VA teacher into two few-step students with different step budgets, using consistency distillation plus a distribution-matching loss on video. At inference, R2 Spec uses them as Draft and Target: the Target verifies the Draft's action proposals, locally repairs some rejections, and otherwise falls back to full Target generation. On RoboTwin 2.0 the system roughly matches the teacher and is about 17 times faster per call, although the paper notes that this ratio is not normalized for hardware. LIBERO and four real-robot tasks are also evaluated.

Critical strengths
  1. F2 Distill mostly builds on Flash-WAM's consistency setup, adding an adapted DMD loss on video, but it works well. At the same step budgets, both students beat Flash-WAM on RoboTwin 2.0 Clean success (91.5 vs. 88.4 and 92.5 vs. 90.2) and on FVD and LPIPS (Table 2), and the full system ends up on par with the teacher. The paper does not say whether the Flash-WAM success rates in Tables 1 and 2 were obtained with the authors' own RoboTwin protocol; Sec. 4.1 only says that "Baseline rows may use other denominators".

  2. The real-robot study is an appropriate test for a method that cuts latency. The four AgileX Piper tasks are dynamic or time-critical, such as catching a block sliding off a box or grasping a rolling ball, and each method gets 100 trials per task. DiSpecWAM averages about 80% against 72.5% for Fast-WAM, the strongest baseline. The paper gives few details on how these runs were set up, for example the training data, the baselines' step budgets or the latency on the robot.

Critical weaknesses
  1. My main concern, and the main reason for my score, is the motivation. The paper claims that generating future video at inference helps control (p. 1, Fig. 1b–c), but never tests this. It even says on p. 2 that "omitting generated future frames does not imply losing learned dynamics or closed-loop control." In DiSpecWAM the video only conditions the actions. Draft video is discarded and the verifier compares actions only, so I consider the benefit of faster video generation unclear. Beating Fast-WAM, a different model that skips test-time video generation, does not resolve this question. The success drop without the FMDD video loss is equally consistent with video prediction acting only as a training-time regularizer. Without the video argument, the difference from Realtime-VLA FLASH stated in Sec. 2 reduces to local repair with a recheck, which adds 0.70 points and saves 51 ms per call (Table 4).

  2. I am also not convinced by the gains credited to R2 Spec. The 205 ms is per inference call, "not compute time per executed action" (Sec. 4.2), since R2 Spec may execute only 8 of 32 actions before replanning. Target-only takes 376 ms per 32-action chunk. If it executes the whole chunk, DiSpecWAM must average more than about 17 executed actions per call to be cheaper per action. The paper does not report this number. Nor can I reconcile the timing. One extra action denoising step costs 27 ms (Table 3), yet the full system is only 22 ms slower than Draft-only. Every call still runs a Target verification. The 1-point gain over Target-only is small and, as the authors acknowledge in Sec. 4.5, confounded with replanning frequency.

Questions for authors

Q1. Beyond conditioning the actions, what is the generated future video used for at inference, and what evidence shows that generating it at test time improves control?

Q2. What do Draft-only and Target-only achieve if they also replan after the first 8 actions, as DiSpecWAM does whenever it accepts only a prefix or a repair?

3 comments

We sincerely thank the reviewer for the sharp, focused evaluation of our test-time video motivation, timing accounting, and controlled comparisons. We address each concern from both architectural principles and controlled empirical evidence below.


W1 & Q1: Role of Test-Time Future Video in Control and Verification

Thank you for raising this foundational question. We clarify both (1) what generated future video actually does inside DiSpecWAM at inference (beyond conditioning Draft actions) and (2) controlled evidence isolating test-time video generation from training-time regularization:

(1) Architectural role: Future video anchors Draft proposals, Target verification, and local repair. First, we apologize that our phrasing on page 2 (“omitting generated future frames does not imply losing learned dynamics…”) read as undercutting our motivation; it was intended only to note that video-skipping policies remain reactive, and we revised Section 1 accordingly. Second, in DiSpecWAM, generated future video is not discarded without affecting verification:

  • Action proposal conditioning: Inside the 30-layer causal Transformer, Draft action queries attend directly to generated future-video latents V^td\hat{V}^d_t to anticipate multi-view contact geometry before committing actions (discarding V^td\hat{V}^d_t after A^td\hat{A}^d_t is produced only means predicted frames are not committed into observation history).

  • Target verification and repair conditioning (Eqs. 9, 12--13): Rather than comparing actions in isolation, the Target verifier's velocity field vTa(A~t,σk,σk,ct−)v^a_T(\tilde{A}_{t,\sigma_k}, \sigma_k, c^-_t) (Eq. 9), scene-shift detector efresh=Distcont(A^T,t,1:Hr−,A^T,t,1:Hrf)e_{\mathrm{fresh}} = \mathrm{Dist}_{\mathrm{cont}}(\hat{A}^-_{T,t,1:H_r}, \hat{A}^f_{T,t,1:H_r}) (Eq. 12), and one-step Euler repair (Eq. 13) are all conditioned on ct−c^-_t and ctfc^f_t---which are precisely the future-video representations generated at test time by the Target video branch. Degrading few-step video directly harms control: removing video FMDD (w/o Distribution Alignment, Table 4) drops Clean SR by 9.33 points (93.50% →\to 84.17%), and removing video consistency while keeping action consistency drops LIBERO SR by 8.20 points (98.45% →\to 90.25%).

(2) Controlled evidence that test-time video generation improves control (ruling out training-only regularization).

  • Same-checkpoint test-time toggle on HumanWAM (100% identical weights): To strictly eliminate training-time regularization as a confounder, we evaluated HumanWAM across RoboTwin 2.0 under Clean, Scene-OOD, and Random protocols using the exact same trained WAM checkpoint with test-time future-video generation bypassed vs. enabled:
Same-Checkpoint Inference ModeClean SR (%)Scene-OOD (%)Random SR (%)
Observation only (test-time video off)73.4448.4423.40
Future video →\rightarrow actions (video on)81.9657.0428.04
Net Test-Time Control Gain+8.52+8.60+4.64
  • Comparison against video-skipping WAMs on dynamic real-world tasks: Fast-WAM and PFD both use future modeling during training (retaining training-time regularization) while skipping iterative video generation at test time. Yet across 400 trials on four dynamic AgileX tasks (including newly added PFD), retaining explicit test-time video prediction in DiSpecWAM reaches 79.75% mean SR vs. 74.00% (PFD) and 72.50% (Fast-WAM), with +9+9 to +12+12-point gains on fast visual-foresight tasks (Grasp Rolling Ball: 83% vs. 72% PFD / 74% Fast-WAM; Grasp Sloped Block: 88% vs. 82% PFD / 76% Fast-WAM).

Distinction from Realtime-VLA FLASH. Beyond retaining calibrated test-time future video via F$^2$Distill, R$^2$Spec decouples expensive multi-view Target video generation from verification via cached ct−c^-_t and efreshe_{\mathrm{fresh}} re-anchoring, and replaces binary fallback with κ\kappa-bounded Euler repair (+0.70+0.70 SR and −51-51 ms over FLASH-style w/o Repair).

W2 & Q2: Matched 8-Step Replanning, Per-Action Cost Break-Even, and Timing Reconciliation

Thank you for this precise quantitative audit. We address matched 8-step replanning, the 17.4-action break-even vs. Target-only, and why total latency is 205 ms (+22 ms over Draft-only vs. +27 ms in Table 3) below:

(1) Draft-only and Target-only with 8-action replanning (Hr=8H_r = 8). We evaluated Draft-only and Target-only when generating and executing 8 actions per call before re-observing and replanning:

Method / VariantSteps (v/a)Exec. PrefixClean SR (%) ↑\uparrowms / call ↓\downarrowms / action ↓\downarrow
Full LingBot-VA (Teacher)25 / 503292.903404.85106.40
Draft-only1 / 23291.48183.005.72
Draft-only (8-step replan)1 / 2891.60176.0022.00
Target-only2 / 43292.50376.0011.75
Target-only (8-step replan)2 / 4893.12354.0044.25
DiSpecWAM (with Repair)1/2 & 2/4Adaptive (8 / 32)93.50205.007.67

RoboTwin 2.0 Clean (50 tasks). The 8-step baselines generate and execute 8 actions per call before re-observing and replanning. Both generation length and replanning frequency change relative to the 32-step baselines.

Generating and executing 8 actions per call yields +0.12+0.12 points for Draft-only (91.60%) and +0.62+0.62 points for Target-only (93.12%, surpassing 32-step open-loop Full's 92.90%), but raises Target-only's cost to 44.25 ms/action. DiSpecWAM reaches 93.50% (+0.38+0.38 points above 8-step Target-only) at 7.67 ms/action (5.77×5.77\times faster), as κ\kappa-bounded repair preserves sharp contact commitment while triggering 8-step replanning selectively upon cross-budget disagreement.

(2) Executed actions per call vs. the 17.4-action break-even point. Your break-even math is exact: since 32-step Target-only costs 376/32=11.75376 / 32 = 11.75 ms/action, DiSpecWAM at 205 ms/call must average over 205/11.75=17.45205 / 11.75 = 17.45 executed actions/call. From Table 5 (δver=0.15,κ=0.75\delta_{\mathrm{ver}}{=}0.15, \kappa{=}0.75), final return routes are 74% Draft (32 actions), 22% Repair (8 actions), and 4% Fallback (32 actions), averaging 0.74×32+0.22×8+0.04×32=26.720.74{\times}32 + 0.22{\times}8 + 0.04{\times}32 = \mathbf{26.72} executed actions/call---well above 17.45, giving 7.67 ms/executed action (1.53×1.53\times faster than 32-step Target-only and 13.87×13.87\times faster than Full LingBot-VA):

Final Return Route (δver=0.15,κ=0.75\delta_{\mathrm{ver}}{=}0.15, \kappa{=}0.75)Call ShareExecuted Actions
Accepted Draft74%32
Accepted Local Repair22%8
Full Target Fallback4%32
Weighted Mean per Call100%26.72
Methodms / call ↓\downarrowActions / callms / action ↓\downarrowSpeedup ↑\uparrow
Full LingBot-VA (25/50, 32-step)3404.8532.00106.401.00$\times$
Target-only (2/4, 8-step replan)354.008.0044.252.40$\times$
Target-only (2/4, 32-step)376.0032.0011.759.06$\times$
Draft-only (1/2, 32-step)183.0032.005.7218.60$\times$
DiSpecWAM (Adaptive 8 / 32)205.0026.727.6713.87$\times$

(3) Timing reconciliation: Why +1 step in Table 3 costs 27 ms, whereas verification + recovery adds only 22 ms (183 →\to 205 ms).

  • Why Table 3's extra step costs 27 ms: Moving from (1, 1) (156 ms) to (1, 2) (183 ms) adds a sequential cross-modal Transformer step, whereas most of the 156 ms base latency comes from multi-view RGB encoding and full-sequence video denoising.

  • Why Verify costs far less than 27 ms: Initial verification is not a sequential cross-modal step. On every call, it reuses already-encoded current observations and cached Target video features ct−c^-_t (skipping Target video generation on all 74% accepted calls) and evaluates both noise levels Sinit={0.25,0.5}S_{\mathrm{init}}{=}\lbrace 0.25, 0.5 \rbrace in one GPU-batched action-branch forward pass, costing only a few milliseconds.

  • Why amortized recovery adds only $\sim$16 ms (saving 51 ms over w/o Repair): Fresh Target context ctfc^f_t runs on only the 26% rejected calls. Without repair (w/o Repair, Table 4), all 26% rejections run full 4-step Target generation, raising mean latency to 256 ms (+73+73 ms over Draft-only). Replacing 22% of those full rollouts with 1-step Euler repair saves 51 ms on average (256→205256 \to 205 ms), reconciling Tables 3--5.

Protocol details. Section 4.1/4.3 also clarifies that Table 2 evaluates Flash-WAM on the identical LingBot-VA backbone, 25k RoboTwin dataset, and 50 tasks×100 trials50\text{ tasks}\times 100\text{ trials} protocol, and reports real-robot fine-tuning details alongside the 400-trial PFD comparison (79.75% vs. 74.00%).

Thank you for the rebuttal and the new per-action and 8-step results. The per-action accounting resolves my question on cost: at 26.72 executed actions per call, DiSpecWAM is at 7.67 ms per action against 11.75 ms for Target-only. On success, however, Target-only with 8-step replanning already reaches 93.12%, so only 0.38 of the 1.0-point gain over Target-only remains.

My main concern about the motivation is not resolved. The same-checkpoint toggle is run on HumanWAM, a model that is not in the paper and not described in the response, rather than on DiSpecWAM. Switching off an input the model was trained to use is also expected to hurt, so this test does not separate the test-time role of the video from its role in training. The Fast-WAM and PFD comparisons still involve different models, and the FMDD ablation changes training, not inference.

I appreciate the additional experiments, but I still consider the benefit of generating video at test time not shown for DiSpecWAM, so I am keeping my score.

Reviewer_Zhh3OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes an acceleration framework for World Action Models (WAMs) to reduce inference latency while retaining future video prediction. They first distill a full-capacity teacher model into a lightweight draft model and a slightly heavier target model using consistency and distribution matching losses. At test time, they adopt a speculative inference scheme where the target model verifies the draft model's action proposals, locally repairs rejected action prefixes, or falls back to full target generation. Experiments on simulation benchmarks and real-world robotic manipulation tasks show that the method significantly speeds up inference while maintaining task success rates comparable to the full model.

Critical strengths
  1. The overall framework is reasonable and well-motivated. Combining two distilled student models with online action verification and local repair provides a practical way to balance inference speed and control reliability while keeping future video generation.
  2. The experimental evaluation is solid and covers both simulation benchmarks (RoboTwin 2.0 and LIBERO) and real-world dynamic manipulation tasks, demonstrating a good trade-off between latency and success rate.
Critical weaknesses
  1. Technical novelty and relation to prior work: The proposed framework appears somewhat combinatorial. The distillation stage mainly combines the modality-specific consistency parameterization (similar to Flash-WAM) with distribution matching distillation (DMD), and the inference stage extends existing speculative decoding schemes with a one-step Euler repair and threshold rules. It would be helpful if the authors could further clarify the unique technical insights beyond combining these existing components.
  2. Assumptions in the verification and repair mechanism:
    • Cached visual context: From my understanding, the initial verification relies on cached target video features from previous steps, and this cache is not updated when the draft proposal is accepted (either full horizon H=32H=32 or partial prefix Hr=8H_r=8). In fast-changing dynamic environments, it seems that the cached visual context could quickly become stale. It would be better to discuss how this affects verification reliability.
    • Shared failure modes and hyperparameters: Since both the draft and target models are distilled from the same teacher with the same objective and only differ slightly in denoising steps, they might share similar failure modes. In addition, the speculative repair pipeline introduces quite a few manual thresholds and noise-level hyperparameters, which might require careful tuning across different setups.
  3. Missing experimental controls and efficiency metrics:
    • Latency per executed action: While the authors note that per-call latency (205 ms) is not the compute time per executed action, the main comparisons (e.g., 16.61× speedup in Table 1 and Figure 1) still rely solely on per-call latency. Since partial execution and repair only execute Hr=8H_r=8 steps instead of H=32H=32 and must replan 4× more frequently, it would be more informative to also report the average compute time per executed action (or total inference time per episode).
    • Replanning frequency in ablations: In Table 4, the full method achieves a higher success rate than the Target-only baseline. As executing a shorter prefix (Hr=8H_r=8) naturally increases the closed-loop replanning frequency, it remains unclear how much of this gain comes from the repair mechanism itself versus more frequent replanning. Adding a baseline with matched replanning frequency (H=8H=8) would make this comparison much more convincing.
    • Unisolated ablations and visualization: Although the paper acknowledges in the text that "w/o Consistency" in Table 4 removes both video and action consistency losses and that Figure 3 reuses the full model checkpoint instead of the distilled student, leaving these unisolated in the main experiments weakens the empirical validation of F²Distill.
  4. Writing and presentation: Several discussions in Section 4 read like disclaimers or responses to previous reviews (explicitly stating what their experiments or figures "cannot isolate" or that baseline numbers come from different protocols) rather than directly refining the experimental setups to make the paper self-contained.
Questions for authors
  1. When the method triggers partial execution or local repair, it only executes a shorter prefix (Hr=8H_r=8) instead of the full horizon (H=32H=32) and replans sooner. Could the authors report the average compute time per executed action (or total inference time per episode) alongside per-call latency? Also, what is the GPU memory overhead of loading both the draft and target models simultaneously?
  2. Following the point above, executing 8 steps leads to more frequent closed-loop replanning than the baselines that execute 32 steps. How much of the performance gain of DiSpecWAM (93.50%) over Target-only (92.50%) in Table 4 comes from the repair mechanism versus simply replanning more frequently? It would be great to see results for Draft-only and Target-only evaluated at H=8H=8.
  3. In Algorithm 1, the cached target context (ct−c_t^-) is not refreshed whenever draft proposals are accepted (Lines 6–9). During consecutive accepted drafts, how many steps can this cache remain un-updated in practice, and why does verifying current actions against an old visual context not hurt performance in dynamic tasks?
  4. To directly resolve the caveats noted in Section 4.2 and 4.4, could the authors provide a clean ablation where action consistency is kept while only video consistency is removed, and update Figure 3 using the actual distilled student checkpoint?
4 comments

We deeply appreciate your technically rigorous review. Below we clarify our core mechanistic insights beyond combining existing primitives, analyze cached-context behavior and shared failure modes, and provide all requested efficiency metrics and controlled ablations.


W1: Unique Technical Insights Beyond Combining Existing Components

While F$^2$Distill and R$^2$Spec build on consistency distillation, DMD, and speculative execution, naively combining unimodal diffusion distillation with standard speculative decoding breaks down in coupled World Action Models (WAMs) due to cross-modal gradient interference inside a shared 30-layer Transformer and the heavy cost of multi-view video imagination. Our core insights lie in two asymmetric cross-modal mechanisms and their co-design:

(1) Asymmetric Cross-Modal Trajectory-and-Distribution Calibration in F$^2$Distill. In interleaved causal WAMs, action denoising explicitly attends to generated future video states, creating a structural compression conflict:

  • Why consistency-only (Flash-WAM) is insufficient: Pairwise consistency LCD\mathcal{L}_{\mathrm{CD}} (Eq. 4) enforces local step invariance under a held-fixed cross-modal state cc, but accumulated residuals at 1--2 video steps induce rollout-level covariate shift in Student video endpoints yθvy^v_\theta. Visual realism does not equate to control fidelity (DiSpecWAM v1/a2 has worse FVD than Flash-WAM v2/a4, 73.82 vs. 66.70, yet +1.24+1.24\\% higher Clean SR), because off-manifold video latents distort downstream action conditioning.

  • Why symmetric full-modality DMD is unstable: Directly applying DMD to low-dimensional, sharp contact actions requires an auxiliary fake-score network ϕ\phi. As analyzed in Appendix A.3 (Eq. A.2, ∥Δ+∥2≤(1−r)∥Δ∥2+r∥e∥2\lVert \Delta^+ \rVert_2 \le (1-r)\lVert \Delta \rVert_2 + r\lVert e \rVert_2), whenever fake-score error ∥e∥2\lVert e \rVert_2 rivals the student--teacher discrepancy on sharp action modes, surrogate gradients fail to contract and inject high-variance interference into the shared Transformer.

  • Our insight: F$^2$Distill decouples the two modalities asymmetrically: it enforces fixed-context bimodal consistency (retaining unbiased direct endpoint supervision fθa=x^0,θaf^a_\theta = \hat{x}^a_{0,\theta} for actions) while applying scale-normalized, elementwise γ\gamma-clipped distribution calibration (LFMDD\mathcal{L}_{\mathrm{FMDD}}, Eqs. 5--7) solely to velocity-implied Student video rollout endpoints (bounding coordinate gradients by ωvγ/∥Mv∥1\omega_v\gamma/\lVert M_v \rVert_1). This aligns the action-conditioning visual manifold without corrupting continuous action learning---cutting the Clean SR gap at v1/a2 from 4.48 points (Flash-WAM) to 1.42 points, whereas removing video FMDD or video consistency drops SR by 9.33 points (Table 4) and 8.20 points (LIBERO ablation).

(2) Context-Decoupled Flow Verification and κ\kappa-Bounded Manifold Repair in R$^2$Spec. Unlike discrete LLM speculation or action-only VLAs (e.g., Realtime-VLA FLASH, which has no future-video branch and uses binary accept-or-discard fallback), R$^2$Spec exploits continuous flow geometry: (i) decoupled video-context caching with shift-aware re-anchoring (efreshe_{\mathrm{fresh}}, Eq. 12), checking Sinit={0.25,0.5}S_{\mathrm{init}}{=}\lbrace 0.25, 0.5 \rbrace in one batched action-branch pass and refreshing Target video context only upon rejection (26% of calls); and (ii) discrete-gated one-step Euler manifold projection (Eq. 13), recognizing that rejected Drafts with agreeing gripper gates (qHr=1q_{H_r}{=}1) already capture the right discrete interaction mode and need only local continuous refinement within radius κ\kappa over Hr=8H_r{=}8 steps. Across RoboTwin 2.0 Clean, 74% of calls accept Draft directly, 22% execute locally repaired prefixes, and only 4% require full Target fallback (Table 5), simultaneously reducing latency (256 →\to 205 ms) and raising success (92.80% →\to 93.50%). Crucially, this velocity check and warm-start Euler step rely directly on F$^2$Distill aligning both Students onto a shared probability-flow field.

W2.1 & Q3: Cached Visual Context (ct−c^-_t) Staleness and Dynamic-Task Reliability

Thank you for this sharp question. Verifying against cached ct−c^-_t does not degrade dynamic-task control due to three built-in fail-safe properties in Algorithm 1:

  • Live observation grounding at every call: Draft always generates actions from the latest observation oto_t, and Verify still conditions on current camera frames, robot proprioception, and executed action history (Sec. 3.3)---only the Target's predicted future-video latents reuse ct−c^-_t.

  • Conservative fail-safe rejection and efreshe_{\mathrm{fresh}} re-anchoring: Staleness errs toward safe rejection, never silent error propagation. In fast-changing scenes, if ct−c^-_t becomes stale, live-conditioned Draft proposals A^td\hat{A}^d_t diverge from stale-conditioned Target velocities (dh(k)>δverd^{(k)}_h > \delta_{\mathrm{ver}}), immediately triggering rejection, forcing a fresh TargetContext(ot)(o_t) update (Line 11), and switching to fresh context ctfc^f_t whenever efresh>δfreshe_{\mathrm{fresh}} > \delta_{\mathrm{fresh}} (Eq. 12, Appendix B.2).

  • Practical refresh horizon and dynamic real-robot validation: With a 74% overall Draft acceptance rate (76--83% on contact-transition tasks below, where 21% of calls refresh ctfc^f_t via 17.5% repair + 3.5% fallback), standard episodes span only a few 32-step chunks, so ct−c^-_t typically bridges 2--4 smooth transport calls between contact-triggered refreshes. Most directly, on four fast-changing dynamic AgileX real-robot tasks---catching a sliding block (88%), catching a falling ball (76%), grasping a rolling ball (83%), and timed bell pressing (72%)---DiSpecWAM achieves 79.75% mean success, outperforming PFD (74.00%), Fast-WAM (72.50%), and LingBot-VA (68.25%).

TaskAccepted Draft (%)Context Refresh (%)Repair (%)Fallback (%)
Hanging Mug7624204
Turn Switch7723194
Place Can Basket8020173
Press Stapler8317143
4-Task / 50-Task Mean79 / 7421 / 2617.5 / 22.03.5 / 4.0

W2.2: Shared Failure Modes and Hyperparameter Robustness

Shared vs. budget-induced failure modes. We distinguish two error sources: (i) step-compression discretization errors, where 1/2-step Draft drifts off the flow path; Target's 2×2\times budget and multi-level velocity check at Sinit={0.25,0.5}S_{\mathrm{init}}{=}\lbrace 0.25, 0.5 \rbrace reliably catch these, explaining the +2.02+2.02-point gain from Draft-only (91.48%) to DiSpecWAM (93.50%); and (ii) inherited Teacher perception/policy blind spots, where Draft and Target share systematic bias from the same Teacher. As noted in Section 4.5 and Appendix B.1, verification bounds deviation from the Target flow rather than guaranteeing task success when the Teacher itself errs, defining the natural boundary of self-distilled speculation.

Zero per-task hyperparameter tuning. Because our noise levels sit on the standardized [0,1][0,1] flow-matching quartile grid, all verification and repair hyperparameters---Sinit={0.25,0.5}S_{\mathrm{init}}{=}\lbrace 0.25, 0.5 \rbrace, (σh,σl,σre)=(0.5,0.25,0.375)(\sigma_h, \sigma_l, \sigma_{\mathrm{re}}){=}(0.5, 0.25, 0.375), δfresh=δver=0.15\delta_{\mathrm{fresh}}{=}\delta_{\mathrm{ver}}{=}0.15, κ=0.75\kappa{=}0.75, and Hr=8H_r{=}8---are held strictly fixed with zero per-task or per-benchmark tuning across all 50 RoboTwin 2.0 Clean/Random tasks, all 4 LIBERO suites (98.9% SR), and all 4 AgileX real-robot tasks (79.75% SR), with neighboring settings in Tables 5--6 maintaining 92.94%--93.50% Clean SR.

W3.1 & Q1: Latency per Executed Action and Dual-Model GPU Memory

Using the exact return-route shares from Table 5 (74% Draft executing 32 steps, 22% Repair executing 8 steps, and 4% Fallback executing 32 steps), DiSpecWAM executes a weighted mean of 26.72 actions per call. At 205.00 ms per call, this gives 7.67 ms per executed action---13.87$\times$ faster than Full LingBot-VA (106.40 ms/action), 1.53$\times$ faster than 32-step Target-only (11.75 ms/action), and 5.77$\times$ faster than 8-step Target-only (44.25 ms/action):

Final Return Route (δver=0.15,κ=0.75\delta_{\mathrm{ver}}{=}0.15, \kappa{=}0.75)Call ShareExecuted Actions
Accepted Draft74%32
Accepted Local Repair22%8
Full Target Fallback4%32
Weighted Mean per Call100%26.72
Methodms / call ↓\downarrowActions / callms / action ↓\downarrowSpeedup ↑\uparrow
Full LingBot-VA (25/50, 32-step)3404.8532.00106.401.00$\times$
Target-only (2/4, 8-step replan)354.008.0044.252.40$\times$
Target-only (2/4, 32-step)376.0032.0011.759.06$\times$
Draft-only (1/2, 32-step)183.0032.005.7218.60$\times$
DiSpecWAM (Adaptive 8 / 32)205.0026.727.6713.87$\times$

GPU memory overhead. Measured on a 48 GB workstation RTX 4090 GPU in bfloat16, peak inference memory is 16.48 GiB for a single Student and 35.38 GiB when loading Draft and Target concurrently (+18.90+18.90 GiB, 2.15×2.15\times), as our implementation instantiates two independent pipelines with separate context buffers; sharing the frozen vision/language encoders can readily reduce this footprint for 24 GB GPUs.


W3.2 & Q2: Isolating Repair Under Matched Replanning Frequency (Hr=8H_r = 8)

Following your suggestion, we evaluated Draft-only and Target-only at an Hr=8H_r{=}8 generation and execution horizon:

Method / VariantSteps (v/a)Exec. PrefixClean SR (%) ↑\uparrowms / call ↓\downarrowms / action ↓\downarrow
Full LingBot-VA (Teacher)25 / 503292.903404.85106.40
Draft-only1 / 23291.48183.005.72
Draft-only (8-step replan)1 / 2891.60176.0022.00
Target-only2 / 43292.50376.0011.75
Target-only (8-step replan)2 / 4893.12354.0044.25
DiSpecWAM (with Repair)1/2 & 2/4Adaptive (8 / 32)93.50205.007.67

RoboTwin 2.0 Clean (50 tasks). The 8-step baselines generate and execute 8 actions per call before re-observing and replanning. Both generation length and replanning frequency change relative to the 32-step baselines.

Generating and executing 8 actions per call yields +0.12+0.12 points on Draft-only (91.60%) and +0.62+0.62 points on Target-only (93.12%), with both shorter generation and more frequent visual feedback, but at a 3.77×3.77\times per-action compute cost (44.25 ms/action). Crucially, DiSpecWAM achieves a further gain to 93.50% (+0.38+0.38 points over 8-step Target-only) at 5.77×5.77\times lower cost per executed action (7.67 ms/action), because κ\kappa-bounded local repair refines Draft actions without multi-step over-smoothing and triggers 8-step replanning selectively only when cross-budget disagreement arises.

W3.3 & Q4: Isolated Video-Consistency Ablation and Figure 3 Checkpoint Clarification

Figure 3 checkpoint clarification. Upon auditing our feature archive (internal_feature_tsne_coordinates.npz), we confirmed that both Figure 3 (v2/a4) and Figure 4 (v1/a2) were generated using our distilled F$^2$Distill Student checkpoints; the caveat in Section 4.2 (line 351) stating that Figure 3 reused Full's checkpoint was introduced erroneously during pre-submission editing due to an ambiguous internal key name (lingbot_v2a4). We removed that erroneous caveat in Section 4.2 and updated Figure 3's legend to Ours (Distilled v2/a4).

Clean isolated video-consistency ablation. On LIBERO, we trained a Target variant that retains both action consistency and video FMDD while removing only video consistency (LCDv\mathcal{L}^v_{\mathrm{CD}}):

Distillation ObjectiveSpatialObjectGoalLongAvg. SR (%) ↑\uparrow
LingBot-VA Teacher (25/50)98.5099.6097.2098.5098.45
Target (Video + Action Cons. + FMDD)98.4099.8098.4097.2098.45
w/o Video Cons. (Action Cons. + FMDD kept)91.2089.3092.3088.2090.25

Removing video consistency alone causes an 8.20-point drop (98.45% →\to 90.25%). Together with Table 4's w/o Distribution Alignment (−9.33-9.33 points without video FMDD), this cleanly proves both video objectives are indispensable.


W4: Self-Contained Presentation in Section 4

In the revision, rather than keeping methodological caveats, we directly integrate per-executed-action latency (Sec. 4.2), the verified distilled-checkpoint caption (Fig. 3), the 400-trial PFD real-robot benchmark (Sec. 4.3), and the Hr=8H_r{=}8 matched-prefix and isolated video-consistency ablations (Sec. 4.4) to make Section 4 fully self-contained.

Reviewer_A69KOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Dispecwam addresses the high computational cost of jointly generating future videos and robot actions in world action models, and the reliability issues associated with few-step generation. F2Distill distills a multi-step teacher into few-step models by using consistency distillation and flow-matching distribution alignment. R2Spec then follows a Draft-Verify-Repair-Reverify procedure. The Draft model generates an action candidate. The Target model checks it by adding noise and recovering the action with its velocity field at different noise levels. Candidates that fail verification are locally repaired and checked again, and if repair fails, the method falls back to full Target generation.

Critical strengths
  1. The paper studies the practical problem of reducing the cost of joint video and action generation while maintaining task performance in world action models, which is valuable for robot control.

  2. The proposed repair strategy is a useful safeguard. When a Draft action is unreliable, the method applies a limited correction, checks the result again, and uses the higher-budget Target model if repair fails. This prevents unverified Draft actions from being executed directly, although its effectiveness still needs stronger empirical support.

Critical weaknesses
  1. The method first distills the full WAM into two Students with F2Distill. As expected, both Students perform slightly worse than the full model when used alone. However, after the R2Spec stage, the system outperforms the full model. This is counterintuitive, because the final system consists only of Students distilled from the full model. The current ablations do not explain where this additional gain comes from, and more analysis is needed to show why R2Spec can surpass the full model.

  2. Apart from the repair mechanism, the technical novelty is limited. F2Distill builds on consistency distillation and adds distribution alignment, and both ideas are well established in diffusion distillation. Few-step distillation of WAMs has also been studied in Flash-WAM, and the Draft-Target framework of R2Spec follows speculative decoding. The paper should clarify what is new beyond adapting these techniques to WAMs.

Questions for authors
  1. Figure 2 omits several conditions in Algorithm 1 and labels accepted repairs as Draft actions. Could the authors clarify the relationship between the two and which procedure is used in the experiments?

  2. In Table 1, all values of DiSpecWAM are bolded, although several of them are not the best. For example, PFD achieves higher RoboTwin Random and Avg. SR, and LingBot-VA achieves higher LIBERO Long SR. What rule was used for the highlighting?

  3. The analyses in Table 2, Figures 3-4, and the real-world evaluation (Figure 6) mainly compare DiSpecWAM with Flash-WAM and Fast-WAM. However, Table 1 shows that MECo-WAM and PFD are also strong baselines, and PFD even achieves a higher RoboTwin average SR than DiSpecWAM. The real-world section also describes Fast-WAM as the strongest baseline, although both methods outperform it in Table 1. Why are these two methods not included, at least in the real-world evaluation?

  4. Could the authors apply the proposed method to distill other WAM backbones and report the results? This would help show whether the proposed distillation method is effective beyond LingBot-VA.

  5. The current ablations do not explain why R2Spec improves performance beyond the Target model. The authors acknowledge in Section 4.5 that different execution-prefix lengths and replanning frequencies prevent attributing the gains to repair, and that matched-prefix controls are needed. Despite this acknowledgment, the counterintuitive result still requires an explanation. Could the authors provide these controls?

  6. In Figure 8, success rates increase as the denoising budget decreases on all four tasks, with the full WAM performing worst and Draft (1/2) outperforming Target (2/4). This is the opposite of the trend in Table 3, where the full-budget model performs best on average. How were these four tasks selected, and could the authors report per-task results for all tasks?

4 comments

We sincerely appreciate the reviewer's careful reading and constructive feedback. Below we trace step by step why our speculative student system surpasses the full-budget Teacher, clarify our mechanistic novelty, and provide all requested controls and baselines.


W1 & Q5: Why R$^2$Spec Surpasses the Full Teacher and Matched-Prefix Controls

Thank you for raising this central question. Surpassing the Full Teacher (93.50% vs. 92.90% Clean SR) seems counterintuitive only if the Full WAM is viewed as a closed-loop control oracle. Physically, however, Full LingBot-VA (25/50) executes all H=32H{=}32 predicted actions open-loop (3404.85 ms/call) and integrates 50 iterative velocity steps, exposing it to (i) 32-step open-loop horizon drift and (ii) multi-step trajectory over-smoothing at narrow-tolerance contact transitions. Our controlled ablations isolate three progressive mechanisms behind the gain:

(1) Flow verification filters few-step outliers while keeping contact decisiveness (91.48% →\rightarrow 92.80%). Across all 50 tasks, distilled Target (2/4) already approaches Full (25/50) (92.50% vs. 92.90%). Yet more denoising steps are not better everywhere: on narrow-tolerance contact tasks (e.g., hooking a mug handle or toggling a switch), 50-step velocity averaging smooths out sharp corrective motions, whereas 1--2 consistency-distilled steps preserve decisive mode commitment (Draft >> Target >> 4/8 >> Full):

TaskFull (25/50)Higher (4/8)Target (2/4)Draft (1/2)DiSpecWAM
Hanging Mug4047566471
Turn Switch4453617278
Place Can Basket8184899497
Press Stapler6874808793
Avg. (all 50 tasks)92.9092.7492.5091.4893.50

In w/o Repair (Table 4), multi-level flow verification retains consistent Draft trajectories and falls back to Target only on unstable proposals (256 ms/call), lifting Clean SR from 91.48% (Draft) and 92.50% (Target) to 92.80%---nearly matching Full (92.90%), though still limited by 32-step open-loop execution.

(2) 8-action generation and replanning lift a single Student above 32-step Full (92.50% →\rightarrow 93.12% >> 92.90%). Following your suggestion, we evaluated Draft-only and Target-only with an Hr=8H_r{=}8 generation and execution horizon:

Method / VariantSteps (v/a)Exec. PrefixClean SR (%) ↑\uparrowms / call ↓\downarrowms / action ↓\downarrow
Full LingBot-VA (Teacher)25 / 503292.903404.85106.40
Draft-only1 / 23291.48183.005.72
Draft-only (8-step replan)1 / 2891.60176.0022.00
Target-only2 / 43292.50376.0011.75
Target-only (8-step replan)2 / 4893.12354.0044.25
DiSpecWAM (with Repair)1/2 & 2/4Adaptive (8 / 32)93.50205.007.67

RoboTwin 2.0 Clean (50 tasks). The 8-step baselines generate and execute 8 actions per call before re-observing and replanning. Both generation length and replanning frequency change relative to the 32-step baselines.

Generating and executing 8 actions per call lifts Target-only by +0.62+0.62 points to 93.12%, already surpassing the 32-step Full Teacher (92.90%) while raising per-action cost from 11.75 to 44.25 ms/action.

(3) κ\kappa-bounded local Euler repair adds gains beyond frequent replanning (93.12% →\rightarrow 93.50%) at 5.77×5.77\times lower cost. Crucially, DiSpecWAM outperforms both 8-step Target-only (93.50% vs. 93.12%) and both individual Students on contact tasks like Hanging Mug (71% vs. 64% Draft, 56% Target). When verification rejects a proposal at contact (Linit=0L_{\mathrm{init}}{=}0) with agreeing gripper gates (qHr=1q_{H_r}{=}1), repair does not overwrite the Draft with an over-smoothed 32-step Target rollout. Instead (Eqs. 12--13), it anchors on A^td\hat{A}^d_t at σh=0.5\sigma_h{=}0.5, applies one κ\kappa-bounded Target Euler step (σh→σl=0.25\sigma_h \to \sigma_l{=}0.25) on the immediate Hr=8H_r{=}8 prefix under fresh context, and re-observes after 8 steps (22% of calls). Thus, R$^2$Spec triggers 8-step closed-loop refinement selectively at cross-budget disagreements, combining Draft's sharp contact commitment with Target's local precision at 7.67 ms/action.

W2: Technical Novelty Beyond Adapting Existing Distillation and Speculative Components

Thank you for prompting us to clarify our methodological novelty. Naively combining unimodal consistency/DMD distillation with standard speculative decoding fails in World Action Models (WAMs) because video and action share a coupled 30-layer Transformer and explicit multi-view video imagination is computationally heavy. Beyond adapting individual primitives, DiSpecWAM introduces two cross-modal mechanisms and their co-design:

(1) Asymmetric Cross-Modal Distillation in F$^2$Distill. In causal WAMs, action denoising attends directly to generated future video states, creating a compression conflict:

  • Why consistency-only (Flash-WAM) degrades control: Pairwise consistency LCD\mathcal{L}_{\mathrm{CD}} (Eq. 4) constrains local step invariance under a held-fixed cross-modal state cc. At 1--2 steps, accumulated residuals cause rollout-level covariate shift in Student video endpoints yθvy^v_\theta, distorting action conditioning. Notably, visual realism metrics do not reflect control utility: in Table 2, DiSpecWAM v1/a2 has worse FVD than Flash-WAM v2/a4 (73.82 vs. 66.70) yet +1.24+1.24\\% higher Clean SR (91.48% vs. 90.24%).

  • Why symmetric full-modality DMD destabilizes actions: Applying DMD to low-dimensional, sharp contact actions requires an online fake-score network ϕ\phi. As shown in Appendix A.3 (Eq. A.2, ∥Δ+∥2≤(1−r)∥Δ∥2+r∥e∥2\lVert \Delta^+ \rVert_2 \le (1-r)\lVert \Delta \rVert_2 + r\lVert e \rVert_2), when fake-score error ∥e∥2\lVert e \rVert_2 rivals the student--teacher gap, surrogate gradients cease to contract and inject high-variance interference into the shared backbone.

  • Our mechanism: F$^2$Distill resolves this asymmetrically: it applies fixed-context bimodal consistency (retaining unbiased direct endpoint supervision fθa=x^0,θaf^a_\theta = \hat{x}^a_{0,\theta} for actions) while applying scale-normalized, γ\gamma-clipped distribution calibration (LFMDD\mathcal{L}_{\mathrm{FMDD}}, Eqs. 5--7) solely to velocity-implied Student video rollout endpoints (bounding coordinate gradients by ωvγ/∥Mv∥1\omega_v\gamma/\lVert M_v \rVert_1). This calibrates the action-conditioning visual manifold without perturbing action gradients:

Method / BudgetLPIPS ↓\downarrowSSIM ↑\uparrowFVD ↓\downarrowClean SR (%) ↑\uparrow
Full LingBot-VA (v25/a50, Teacher)0.09710.890239.8192.90
Flash-WAM (v1/a2)0.14240.888495.5388.42
DiSpecWAM Draft (v1/a2)0.12870.889273.8291.48
Flash-WAM (v2/a4)0.12480.888966.7090.24
DiSpecWAM Target (v2/a4)0.10960.889749.6492.50

Bold indicates the best few-step result within each matched denoising budget (v1/a2 and v2/a4).

(2) Context-Decoupled Flow Verification and Manifold Repair in R$^2$Spec. Unlike discrete LLM speculation (token likelihood ratios and KV-prefix reuse) or action-only Realtime-VLA FLASH (no video branch and binary rejection fallback), R$^2$Spec exploits continuous flow geometry via: (i) decoupled Target video-context caching with scene-shift re-anchoring (efreshe_{\mathrm{fresh}}, Eq. 12), verifying Sinit={0.25,0.5}S_{\mathrm{init}}{=}\lbrace 0.25, 0.5 \rbrace in one batched action pass while refreshing video context only on rejections (26% of calls); and (ii) discrete-gated, κ\kappa-bounded Euler repair (Eq. 13), locally refining 22% of calls and leaving only 4% to full Target fallback---cutting latency (256 →\to 205 ms) while raising SR (92.80% →\to 93.50%). Crucially, R$^2$Spec's velocity check and warm-start repair rely directly on F$^2$Distill aligning both Students onto a shared flow trajectory (removing FMDD drops SR by 9.33 points, Table 4).


Q1: Alignment of Figure 2 with Algorithm 1

All experiments strictly execute Algorithm 1, whereas Figure 2 was a simplified schematic. Repair is attempted only when Linit=0L_{\mathrm{init}}{=}0 and gripper gates agree (qHr=1q_{H_r}{=}1), and executed only if the repaired prefix passes the independent recheck (Lre≥Hr,qre=1L_{\mathrm{re}} \ge H_r, q_{\mathrm{re}}{=}1) and bound ∥Mr⊙(A^tr−A^td)∥F≤κ\lVert M_r \odot (\hat{A}^r_t - \hat{A}^d_t) \rVert_F \le \kappa; cases with 0<Linit<Hr0 \lt L_{\mathrm{init}} \lt H_r proceed directly to FullTarget (Lines 19--20). In revised Figure 2, we relabel accepted repairs from “Draft actions” to “Repaired actions” and update “Accept >0> 0” to “Accept ≥Hr\ge H_r (88 or 3232)”.


Q2: Highlighting Convention in Table 1

We apologize for the confusion. We originally shaded and bolded the DiSpecWAM row to make our method easy to locate, not to claim column-best values everywhere. We have revised Table 1 and all other tables so that boldface strictly marks the best value in each column (or matched budget group).

Q3: Additional Baselines (MECo-WAM and PFD) in Analyses and Real-World Evaluation

Table 2 and Figure 3 analyze explicit future-video fidelity and video-token features, so we compare against Flash-WAM---the only few-step baseline that explicitly generates future video at test time---while Figure 4 pairs this with trajectory tracking at the matched v1/a2 budget (Fast-WAM and PFD omit test-time few-step video generation). For real-world evaluation, MECo-WAM has no publicly released code or checkpoints. Following your advice, we evaluated PFD across 400 new real-robot trials on the same four AgileX Piper tasks (100 trials/task) and updated Section 4.3 accordingly:

MethodSloped BlockCatch Ball BowlRolling BallPress Bell TimedAvg. SR (%) ↑\uparrow
LingBot-VA7363667168.25
Flash-WAM6875706569.50
Fast-WAM7671746972.50
PFD8267727574.00
DiSpecWAM8876837279.75

DiSpecWAM achieves the highest mean success (79.75% vs. 74.00% for PFD, +5.75+5.75 points), winning all three dynamic catching/grasping tasks by +6+6 to +11+11 points while PFD leads on Press Bell Timed (75% vs. 72%).


Q4: Generalization to Other WAM Backbones (EasyWAM)

We extended the full DiSpecWAM framework (both distillation and speculative inference) to a second WAM backbone, EasyWAM, across all four LIBERO suites on RTX 4090. Compared with the undistilled 10-step EasyWAM Base (Teacher), DiSpecWAM at 4-step and 2-step budgets achieves substantial acceleration with minimal accuracy loss:

BackboneModel / BudgetSpatialObjectGoalLongAvg. (%) ↑\uparrowLatency (ms) ↓\downarrow
LingBot-VABase Teacher (25/50)98.599.697.298.598.53404.85
LingBot-VA+ DiSpecWAM99.0100.098.897.898.9205.00
EasyWAMBase Teacher (10 steps)98.298.097.696.897.7380.89 (1.00$\times$)
EasyWAM+ DiSpecWAM (4 steps)97.898.097.596.497.4166.53 (2.29$\times$)
EasyWAM+ DiSpecWAM (2 steps)97.697.897.296.197.291.91 (4.14$\times$)

Bold indicates the best value within each backbone group.

Applying DiSpecWAM to EasyWAM at 4 and 2 steps yields 2.29$\times$ and 4.14$\times$ speedups (down to 91.91 ms) while preserving 97.4% and 97.2% average SR (within 0.3--0.5 points of the 97.7% undistilled Teacher), confirming full-framework generality across distinct WAM backbones.

Q6: Task Selection in Figure 8 and Complete 50-Task Results

Figure 8 selects all four tasks among the 50 RoboTwin 2.0 Clean tasks where Draft (1/2) beats both Target (2/4) and Full (25/50) (†\dagger below; clarified in Figure 8's caption). On the other 46 broader-tolerance tasks, Full and Target operate near ceiling (95.91% and 94.33% vs. 92.54% Draft), matching Table 3's average trend; on the four narrow-tolerance contact tasks, 50-step smoothing and 32-step open-loop drift degrade Full (40%--81%), whereas DiSpecWAM reaches 71%, 78%, 97%, 93% to lift the 50-task mean to 93.50%:

TaskDraft (1/2)Target (2/4)Full (25/50)TaskDraft (1/2)Target (2/4)Full (25/50)
Adjust Bottle9899100Place Can Basket †\dagger948981
Beat Block Hammer969798Place Cans Plasticbox899294
Blocks Ranking RGB879093Place Container Plate979899
Blocks Ranking Size868992Place Dual Shoes889193
Click Alarmclock9899100Place Empty Cup939597
Click Bell9899100Place Fan909294
Dump Bin Bigbin969798Place Mouse Pad939496
Grab Roller9899100Place Object Basket919495
Handover Block889193Place Object Scale949596
Handover Mic899193Place Object Stand939496
Hanging Mug †\dagger645640Place Phone Stand959798
Lift Pot939495Place Shoe929495
Move Can Pot889194Press Stapler †\dagger878068
Move Pillbottle Pad949596Put Bottles Dustbin878992
Move Playingcard Away979899Put Object Cabinet879092
Move Stapler Pad879092Rotate QRcode929496
Open Laptop969798Scan Object919394
Open Microwave889093Shake Bottle9899100
Pick Diverse Bottles919395Shake Bottle Horizontally9899100
Pick Dual Bottles949596Stack Blocks Three899294
Place A2B Left959697Stack Blocks Two929495
Place A2B Right949697Stack Bowls Three909294
Place Bread Basket939596Stack Bowls Two949597
Place Bread Skillet939596Stamp Seal909295
Place Burger Fries979899Turn Switch †\dagger726144

50-task mean: Draft 91.48%, Target 92.50%, Full 92.90%. DiSpecWAM 50-task mean: 93.50%. †: 71, 97, 93, 78 (DiSpecWAM). † marks the four Figure 8 tasks.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.