acceptodds
Under review as a conference paper at ICLR 2027

MoS-VLA: A Vision-Language-Action Model with One-Shot Skill Adaptation

Abstract

Vision-Language-Action (VLA) models trained on large-scale robotics datasets promise robust control across diverse domains and embodiments, yet they frequently fail zero-shot in novel scenarios. We introduce Mixture-of-Skills VLA (MoS-VLA), a framework that formulates robot manipulation policies as continuous linear combinations of learned basis functions. Pretrained jointly on the Open X-Embodiment data mixture, MoS-VLA constructs a skill space for rapid adaptation. At test time, adaptation in our evaluated robot settings uses a single expert demonstration. We infer the optimal mixture coefficients via a lightweight, gradient-free convex optimization that minimizes L1 action error. MoS-VLA achieves 70-100% success in simulated and real-robot tasks, where baseline success is 0% in all but one comparison. It also achieves lower action-prediction error than two pretrained VLAs and two in-context baselines on their evaluated datasets, while adapting at a fraction of the compute of gradient-based finetuning.

Primary area:
applications to robotics, autonomy, planning
Keywords:
VLA, few-shot, adaptation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_US52OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

MoS-VLA learns 16 image-and-language-conditioned basis functions on an OpenVLA backbone. It adapts to a new context by fitting their coefficients to demonstration actions through convex L1 regression, with no model weight updates. The paper evaluates action prediction, adaptation cost, and robot control. Appendix H extends the method to latent action chunks.

Critical strengths

The adaptation step is simple and practical. Table 2 shows much lower adaptation time and GPU memory use than gradient-based finetuning, with competitive prediction error at 128 examples.

The paper includes simulation and real-robot results. The calibrated single-head control in Figure 4 is also useful: the multiple-basis model lowers offline error on eight of ten held-out datasets.

Critical weaknesses
  1. The main robot comparison lacks matched adaptation controls. Table 1 includes zero-shot OpenVLA and pi0.5, plus ICRT and X-VLA, but no one-demonstration OFT/LoRA baseline or calibrated single-head control. These controls appear only in offline evaluations. The results therefore leave open whether the proposed adaptation method offers a better tradeoff between adaptation cost and closed-loop success.

  2. One-shot reliability is not tested across independent calibration demonstrations. Repeated rollouts of one calibrated policy do not show how sensitive adaptation is to the chosen demonstration. Tables 2 and 6 vary sample counts, not trajectory selection or state coverage. The limited task variation and disabled forward/backward motion in three tasks further narrow the evidence for generalization.

  3. Figure 6 shows uneven approximation quality across unseen domains, but does not link projection residuals to control performance. It remains unclear whether a high residual identifies contexts where coefficient fitting will fail, or when additional demonstrations can help. This matters because adaptation cannot change the learned function span.

  4. The latent extension provides limited evidence for the added basis. At 128 examples, calibration alone gives L1 error 0.369, while the 0.184 result also finetunes the VAE. The authors clearly distinguish these settings. However, the matched VAE without basis functions reaches 0.189 (Table 7), leaving only a small average gain without uncertainty estimates. The stronger latent result therefore does not establish a gradient-free improvement.

Questions for authors
  1. Can the calibrated single-head model and one-demonstration OFT/LoRA be included in the robot comparison?
  2. How were calibration trajectories selected, and which initial states or scene conditions change during testing?
  3. Are fitted coefficients shared across language tasks in the same lab? Are finetuned VAE weights shared across target contexts?
1 comment

Thank you for the detailed review and for highlighting our offline efficiency and k=1k=1 gains. We have now addressed the closed-loop concern with LIBERO-Plus (200 trials/cell, Appendix G) and Franka block-lifting baselines, where MoS-VLA reaches 10/10 in under 10s while calibrated k=1k=1 and 1-demo OpenVLA-OFT both score 0/10.


Q1. Calibrated single-head (k=1k=1) and 1-demo OFT/LoRA in robot comparisons

MoS-VLA outperforms k=1k=1 in LIBERO-Plus and beats both k=1k=1 (0/10) and 1-demo OpenVLA-OFT (0/10) on real-robot block lifting:

  • LIBERO-Plus (Appendix G, Table 5): Across 9 contexts (200 trials/cell), MoS-VLA beats calibrated k=1k=1 in-distribution (0.47 vs. 0.32) and on held-out compositions (0.37-0.53 vs. 0.14-0.27).

  • Real-robot block lifting: MoS-VLA achieves 10/10 in under 10s; calibrated k=1k=1 and 1-demo OpenVLA-OFT (500 steps, about 15 min) both get 0/10, while 1-demo LoRA finetuning of π0.5\pi_{0.5} reaches 10/10 only after 500 gradient steps (about 15 min).


Q2. Calibration trajectory selection and test-time initial-state variations

  • Demo-selection sensitivity: Calibration uses one uncurated demonstration per task. Varying which training episode is used within each of the 10 OOD datasets yields an average within-dataset std of only 0.027 around the 0.265 mean L1L^1 (0.245±0.0170.245 \pm 0.017 with ∣α∣≤1|\alpha| \le 1). Using the first episode (as in the main paper, 0.274 L1L^1) is slightly worse than the average across episodes (0.265), confirming we did not cherry-pick a favorable demo.

  • Initial-state and scene variations: In LIBERO-Plus, calibration and evaluation episodes are sampled independently from LIBERO's initial-state distribution across 200 trials/cell and 9 scene/viewpoint shifts.


Q3 & W3-W4. Coefficient/VAE sharing, projection residuals (Fig. 6), and latent basis (App. I)

  • Coefficient sharing:

    • Real robot (task-specific): Each language task uses its own αc\alpha^c. On the real robot, coefficients from reach do not transfer to block lifting (and vice versa) due to limited language grounding in current VLAs.
    • Within the same robot (OOD datasets): On multi-task OOD datasets, αc\alpha^c transfers across different tasks of the same robot. Using another episode's αc\alpha^c achieves 0.253 L1L^1 (vs. 0.308 for predicting zero actions, and 0.222 when fitting αc\alpha^c on 128 frames sampled across multiple episodes).
    • Across different robots: Fitted αc\alpha^c vectors are not shared across different robot embodiments. The mean cosine similarity between αc\alpha^c vectors from different robots is −0.01±0.29-0.01 \pm 0.29 (nearly orthogonal), reflecting distinct kinematics and action scales.
  • VAE weights (Appendix I):

    • Sharing: VAE weights are shared when frozen (Table 7) and finetuned per target dataset in Tables 6 and 8.
    • Latent basis trade-off: At matched cost (Table 8), the latent basis preserves accuracy (0.184 vs. 0.189, with larger gains on CMU Franka 0.185 vs. 0.240 and DLR Pour 0.135 vs. 0.185) while keeping the context representation low-dimensional (k=16k=16).
  • Projection residuals (Fig. 6): Figure 6 measures whether a target task can be expressed by combining the 16 learned basis functions at all, and LIBERO-Plus confirms how this predicts control performance:

    • Low residual (expressible task): Coefficient fitting works and extra demos reduce noise. On held-out LIBERO-Plus compositions of familiar camera and object shifts, MoS-VLA (k=16k=16) reaches 0.37-0.53 success vs. 0.14-0.27 for the calibrated k=1k=1 baseline, and fitting new coefficients beats reusing training coefficients (0.13-0.18 vs. 0.25-0.36 demo L1L^1).
    • High residual (unexpressible task): Linear coefficient fitting cannot recover missing representations. When the camera shifts along unseen axes (rotation with vertical shift), both models fail (≤0.07\le 0.07).
Reviewer_cuNHOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes MoS-VLA, a vision-language-action framework that integrates function encoders into an OpenVLA backbone. Instead of finetuning the whole policy for a new lab setting, MoS-VLA learns kk neural basis action heads that span a policy space. For a new context, it fits a small coefficient vector by$ L1L^1 $least-absolute-error regression, which is solved as a linear program, requiring only forward passes and no gradient updates. The authors train on OXE/Open-X Magic Soup Plus data and evaluate on held-out OOD datasets, simulated Robosuite tasks, and real Franka tasks. They report lower OOD action error than OpenVLA, $\pi_{0.5}$, X-VLA, and ICRT, much lower adaptation compute/memory than LoRA finetuning, and 70--100% success on several short-horizon tasks after calibration with one expert trajectory.

Critical strengths

1. Practical low-overhead adaptation. The main strength is the adaptation mechanism: one expert trajectory, forward passes plus a convex linear program, no backpropagation, and inference cost independent of demonstration length after calibration. This is consequential because gradient-based VLA finetuning is expensive and often impractical in hardware-constrained lab settings. If the empirical claims hold, this is a useful and simple deployment recipe.

2. Strong OOD and real-robot results with one demonstration. The reported gains over zero-shot OpenVLA/π0.5\pi_{0.5} and over X-VLA/ICRT on several tasks are large, especially on real hardware. These results, if robust, support the paper's central claim that a learned basis-function policy space can provide rapid context adaptation.

Critical weaknesses

1. Unclear attribution and insufficient compute-matched baselines. MoS-VLA is initialized from OpenVLA and further trained for 100k gradient steps on 32 GH200 nodes, while the main OpenVLA baseline is zero-shot. The appendix argues this is only ∼6%\sim 6\% of OpenVLA's training budget, but that does not replace a compute-matched control. A continued-training OpenVLA baseline with the same data, optimizer budget, and LoRA/frozen-backbone setup is needed to isolate the effect of the basis-function/function-encoder design. Similarly, the robot-success comparison should include simple adaptation baselines such as last-layer finetuning, LoRA, and OpenVLA-OFT calibrated with the same one trajectory and comparable compute. Without this, the strong success numbers may partly reflect extra training rather than the proposed calibration mechanism.

2. Limited statistical evidence and scope. Real-robot results use m=10m=10 trials per task and simulation uses m=20m=20, with no confidence intervals, multiple seeds, or failure analysis. The tasks are short-horizon and simplified (e.g., forward/backward motion disabled, single front RGB view, no proprioception). The ablations are also single-seed. This makes the 70-100% success claims difficult to assess statistically. In addition, the paper should clarify the OOD split: are all ten held-out datasets/lab settings truly unseen during MoS-VLA training, and how many distinct contexts/tasks are represented? There is also a discrepancy: the main text says 128 transitions are used consistently, while Appendix Table 5 says X-VLA/ICRT use 512 adaptation examples. This needs reconciliation.

Questions for authors
  1. Please define the OOD split precisely. Which datasets/lab settings are held out from MoS-VLA training, and are they also held out from OpenVLA/π0.5\pi_{0.5}/X-VLA/ICRT training? How many distinct contexts and tasks are in the ten OOD datasets?

  2. Can you provide a compute-matched control: continue training OpenVLA from the same checkpoint for the same 100k steps, with the same data mixture and LoRA/frozen-backbone setup, and evaluate it zero-shot and after the same one-trajectory calibration? This would directly test whether the gain comes from the basis-function design rather than additional training.

  3. How are actions normalized for a new lab/embodiment? The paper says training-split statistics are used. For a genuinely new setting, are these statistics valid? How sensitive is calibration to action-scale mismatch, absolute vs relative action assumptions, and camera/viewpoint changes?

  4. During training, each dataset has a 512-pair calibration buffer and coefficients are frozen every 16 steps. How sensitive is the method to this buffer size and calibration frequency? Does the model rely on dataset identity during training, and how does that affect OOD calibration with a single trajectory?

1 comment

Thank you for the positive assessment of our low-overhead adaptation recipe and OOD gains. Our new 100k-step continued-OpenVLA control confirms that extra pretraining does not improve OOD transfer (0.439 vs. 0.225 L1L^1), while LIBERO-Plus adds 200 trials/cell with Wilson 95% CIs and an explicit failure analysis on unseen camera shifts.


Q1 & W2. OOD split definition, statistical power, and 128 vs. 512 reconciliation

None of the 10 evaluation datasets in Figure 4 appear in the pretraining mixtures of MoS-VLA or any open-data baseline (π0.5\pi_{0.5}'s data is closed-source). Together they span 10 lab contexts and at least 83 distinct language instructions.

  • Instruction counts: ManiSkill (≥8\ge 8), RoboCook (≥9\ge 9), CMU PlayFusion (≥41\ge 41), Imperial Sawyer (17), CMU Franka (3), and 1 each on the remaining five. Offline evaluation tests on 10 held-out validation episodes per dataset, while calibration uses only training-split episodes (zero overlap).

  • Statistical power and failure analysis: Closed-loop LIBERO-Plus (Appendix G, Table 5) evaluates 200 trials/cell with Wilson 95% CIs, and sensitivity checks cover all training episodes per OOD dataset (within-dataset std 0.027 around 0.265 L1L^1). We also isolate the failure boundary: MoS-VLA doubles k=1k=1 on held-out compositions within seen camera axes (0.37-0.53 vs. 0.14-0.27), while unseen camera axes (rotation with vertical shift) defeat both models (≤0.07\le 0.07).

  • 128 vs. 512 reconciliation: We updated the Figure 4 caption to state that MoS-VLA and OFT (k=1k=1) use 128 transitions across all datasets, whereas X-VLA uses 512 examples and ICRT uses one trajectory.


Q2. Compute-matched control: continued OpenVLA (100k steps)

Continuing OpenVLA on the same OXE mixture for 100k steps does not improve OOD transfer (0.439 vs. 0.432 stock, compared to 0.225 for MoS-VLA at 128 transitions). On the 5 training datasets, continued training improves L1L^1 from 0.205 to 0.142, confirming that the optimization succeeded on the training mixture without transferring OOD.

ModelOOD-10 Mean L1L^1
OpenVLA (stock)0.432
OpenVLA continued, 100k steps0.439
OFT k=1k=1 control (zero-shot / 1-demo / 128 frames)0.342 / 0.310 / 0.304
MoS-VLA (1-demo / 128 frames)0.274 / 0.225

In closed-loop LIBERO-Plus, k=1k=1 also trails MoS-VLA across all 9 contexts (0.14-0.27 vs. 0.37-0.53 on held-out compositions). On real Franka block lifting, calibrated k=1k=1 and 1-demo OpenVLA-OFT (500 steps, about 15 min) both score 0/10, whereas MoS-VLA scores 10/10 in under 10s.


Q3. Action normalization, action space, and viewpoint changes

Single-demo normalization matches full-dataset normalization (0.263 vs. 0.265 L1L^1), and MoS-VLA adapts within the visual backbone's training support.

  • Normalization and scale: Actions are conventionally normalized to the 1st-99th percentiles of the dataset (OpenVLA, π0\pi_0). In a new lab without dataset statistics, we normalize using the single demo's percentiles. Even though one demo covers only 30-70% of the dataset range, ∑i=1kαigi(s)\sum_{i=1}^k \alpha_i g_i(s) absorbs per-dimension affine rescaling.

  • Action space and viewpoint: MoS-VLA uses relative EEF deltas (unlike X-VLA and ICRT, which assume absolute poses). In LIBERO-Plus (Appendix G), held-out yaw +20∘+20^\circ reduces success from 0.47 to 0.32 (vs. 0.23 for k=1k=1), while camera shifts along unseen axes (rotation with vertical shift) defeat both models (≤0.07\le 0.07).


Q4. Buffer size, calibration interval, and dataset identity

  • Dataset identity: The network never sees a dataset ID at training or test time. During pretraining, dataset indices only route batches to node-local buffers; at test time, αc\alpha^c is solved from the single demo (reusing training coefficients on LIBERO-Plus gives worse calibration-demo L1L^1 of 0.25-0.36 vs. 0.13-0.18).

  • Train-time buffer size and calibration interval: MoS-VLA is insensitive to both hyperparameters. Across 5,000-step runs (seed 7) varying calibration interval {8,16,64}\{8, 16, 64\} and buffer size {64,512}\{64, 512\}, OOD-10 L1L^1 stays within 0.008 (0.267-0.274 at 1 demo; 0.219-0.227 at 128 frames), with the paper setting (interval 16, buffer 512) performing best. Test-time buffer sizes {64,128,256,512}\{64, 128, 256, 512\} are ablated in Figure 9 (Appendix E).

Train-Time Setting (5,000 steps)OOD-10 1-Demo LP (L1L^1)OOD-10 128 Frames (L1L^1)
Interval 16, buffer 512 (paper setting)0.2670.219
Interval 8, buffer 5120.2740.226
Interval 64, buffer 5120.2710.226
Interval 16, buffer 640.2710.227
Reviewer_K8cTOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes MoS-VLA, a vision-language-action model that represents policies as linear combinations of learned basis functions (“mixture of skills”) and adapts to new contexts by solving a small, gradient-free least-absolute-error (L1) linear program to infer mixture coefficients from a single expert trajectory. Built on an OpenVLA backbone with k parallel continuous-action heads, the method modifies the function-encoder training algorithm to operate in an L1 Banach-space setting and introduces a scalable calibration-buffer training recipe. Empirically, MoS-VLA reduces action-prediction error across ten out-of-distribution datasets and achieves 70–100% success on simulated and real-robot tabletop tasks after one-shot calibration, with lower adaptation time and VRAM than gradient-based fine-tuning or attention-based in-context methods.

Critical strengths
  1. Clear problem setup, algorithm description, and architectural choices. Figures show the method’s calibration workflow and empirical trends well.

  2. The adaptation procedure is simple and has a clear practical benefit. Fitting a small coefficient vector avoids backpropagation through the VLA at deployment. Table 2 quantifies the resulting savings: 2.8 seconds and 17.1 GB for 128 calibration examples, with action error comparable to the fine-tuning baselines.

Critical weaknesses
  1. The central comparison is missing from the robot experiments. Table 1 compares MoS-VLA with zero-shot OpenVLA and pi0.5, and with ICRT and X-VLA given five demonstrations. These are useful comparisons, but the calibrated single-head control from Figure 4 and the LoRA or OFT adaptations from Table 2 are not evaluated for closed-loop success. OpenVLA-OFT is particularly relevant because it also uses continuous action prediction. Without these controls using the same demonstration and comparable offline training, it remains unclear how much of the robot success comes from the learned basis rather than ordinary adaptation.

  2. The scope of the evidence is quite narrow. There are only ten trials per physical task and twenty per main simulation task. And forward/backward motion is disabled in three settings, reaching uses a fixed goal, and pen insertion starts with the pen already grasped. The UR5E experiment with a changed camera view is useful, but sensitivity to the chosen demonstration and initial conditions remains unclear. Since the method builds on established function encoders, convincing deployment evidence is especially important. The present results support rapid calibration in these settings, but leave the breadth of one-shot skill adaptation uncertain.

Questions for authors
  1. Closed-Loop Matched-Compute Baselines: What is the closed-loop success rate of a standard OpenVLA (or OpenVLA-OFT) model when finetuned on the exact same 1 demonstration via LoRA (or when only the final projection layer is adapted)? Does the 1-head (k=1k=1) calibrated OFT model achieve non-zero success on the Franka or Robosuite tasks?

  2. Robustness to Suboptimal Demonstrations: The calibration optimization minimizes empirical L1L_1 error against the single provided demonstration. If the human demonstrator exhibits sub-optimal jitter, pauses, or non-straight trajectories, does the linear program overfit to this noise, leading to degraded deployment rollouts?

1 comment

Thank you for the constructive review and for noting the clarity and efficiency of our adaptation step. Our new LIBERO-Plus study (200 trials/cell across randomized initial states) and Franka block-lifting controls directly resolve the missing closed-loop comparisons: on the same OpenVLA backbone, calibrated k=1k=1 and 1-demo OpenVLA-OFT both score 0/10, whereas MoS-VLA scores 10/10 in under 10s.


W1/W2 & Q1. Matched-compute baselines (k=1k=1, 1-demo LoRA/OFT), scope, and sensitivity

Both the 200-trial/cell LIBERO-Plus benchmark and our new real-robot controls show that MoS-VLA's gains come from the k=16k=16 basis instead of extra compute or ordinary 1-demo adaptation.

  • LIBERO-Plus (Appendix G, Table 5): Across 9 contexts (200 trials/cell), MoS-VLA (k=16k=16) beats k=1k=1 everywhere and doubles success on held-out compositions (0.37-0.53 vs. 0.14-0.27).

  • Real-robot block lifting: Under identical 1-demo calibration, k=1k=1 gets 0/10 and 1-demo OpenVLA-OFT (500 steps, about 15 min) gets 0/10, while MoS-VLA gets 10/10 in under 10s. 1-demo LoRA finetuning of π0.5\pi_{0.5} also reaches 10/10, but requires 500 gradient steps (about 15 min).

  • Constrained axes: Forward/backward motion was disabled in three tasks only because all evaluated models receive a single front RGB image without wrist cameras or proprioception.

  • Initial-state sensitivity: In LIBERO-Plus, each cell evaluates 200 trials over randomized initial states, showing the gains hold across varied starting configurations.

  • Demo-selection sensitivity: Varying which training episode is used for calibration within each of the 10 OOD datasets (8-64 episodes each) yields an average within-dataset std of only 0.027 around the 0.265 mean L1L^1 (0.245±0.0170.245 \pm 0.017 with ∣α∣≤1|\alpha| \le 1). Using the first episode (as in the main paper, 0.274 L1L^1) is slightly worse than the average across episodes (0.265), confirming we did not cherry-pick a favorable demo.


Q2. Robustness to suboptimal demonstrations

With ∣α∣≤1|\alpha| \le 1, calibration error stays within 0.02 L1L^1 of the unperturbed demo (0.245) under both action noise and timing defects. Because calibration fits only k=16k=16 scalars over frozen basis functions, it cannot fit frame-level jitter:

Perturbation to Calibration DemoUnconstrained LP (L1L^1)Box-Constrained LP, α∈[−1,1]k\alpha \in [-1, 1]^k (L1L^1)
None (unperturbed)0.2650.245
Action noise σ=0.05/0.1/0.2\sigma = 0.05 / 0.1 / 0.2 (range [−1,1][-1, 1])0.276 / 0.289 / 0.3140.253 / 0.257 / 0.265
Action lag (1 / 2 steps)0.262 / 0.2630.245 / 0.246
Pauses (20% zero arm)0.2620.244
Slow (0.5x arm)0.2710.262
Subsample (50% frames dropped)0.2840.250
Truncated (first 50% / first 25% of demo)0.310 / 0.3750.276 / 0.306

Timing defects cost ≤0.02\le 0.02. Only severe truncation (missing task coverage) hurts substantially.

Comments on the paper

1 comment

We thank the reviewers for their feedback. Our new experiments rule out extra pretraining compute and single-head scaling as the source of MoS-VLA's gains. We also updated the paper with a new LIBERO-Plus evaluation (Appendix G, Table 5, Figure 12) and a clarified Figure 4 caption:


  1. LIBERO-Plus evaluation (Appendix G, Table 5, Figure 12): MoS-VLA (k=16k=16) outperforms a compute-matched single-head control (k=1k=1) in every setting and doubles its success on held-out compositions (0.37-0.53 vs. 0.14-0.27). Both models were continued on the four LIBERO suites plus six single-factor shifts per suite (24 variants total: yaw ±10∘\pm 10^\circ, pitch +10∘+10^\circ, objects 10-12) and tested on libero_spatial and libero_goal (200 trials/cell, Wilson 95% CIs) with identical 1-demo calibration.
ContextSplitMoS-VLA (k=16k=16)Single Head (k=1k=1)
BaseTrain0.47 [0.40, 0.53]0.32 [0.26, 0.39]
Yaw +10∘+10^\circTrain0.41 [0.34, 0.47]0.38 [0.32, 0.45]
Objects 10Train0.58 [0.51, 0.65]0.45 [0.38, 0.51]
Yaw +20∘+20^\circHeld-out value0.32 [0.26, 0.39]0.23 [0.18, 0.29]
Objects 13Held-out value0.46 [0.39, 0.52]0.35 [0.29, 0.42]
Objects 13 + Yaw 10∘10^\circHeld-out comp.0.53 [0.46, 0.60]0.27 [0.21, 0.34]
Objects 11 + Yaw 20∘20^\circHeld-out comp.0.37 [0.31, 0.44]0.14 [0.10, 0.20]
Objects 13 + Yaw 20∘20^\circHeld-out comp.0.43 [0.37, 0.50]0.20 [0.15, 0.26]
Cam. rot. + vertical shiftUnseen axes0.07 [0.05, 0.12]0.02 [0.01, 0.05]

  1. Compute-matched pretraining controls: Extra training on OXE does not explain our gains. Continuing discrete OpenVLA for 50k or 100k steps stays at 0.443 and 0.439 OOD L1L^1 (vs. 0.407 stock and 0.225 for MoS-VLA). Because the released OpenVLA checkpoint finished Stage 2 on the narrower oxe_magic_soup_plus_minus mixture (which drops DROID and Language Table), continuing on oxe_magic_soup_plus slightly shifts its OOD predictions while keeping training-split L1L^1 flat (0.139-0.142).

  1. Real-robot controls (Franka block lifting): MoS-VLA adapts to 10/10 in under 10s, whereas calibrated k=1k=1 and 1-demo OpenVLA-OFT finetuning (500 steps, about 15 min) both get 0/10. 1-demo LoRA finetuning of π0.5\pi_{0.5} reaches 10/10 only after 500 gradient steps (about 15 min).

  1. 1-demo reliability under realistic test-time conditions: 1-demo calibration works reliably without curating a clean demo or knowing dataset statistics beforehand:
    • Insensitive to demo choice: Varying which training episode is used within each of the 10 OOD datasets yields an average within-dataset std of only 0.027 around the 0.265 mean L1L^1 (0.245±0.0170.245 \pm 0.017 with ∣α∣≤1|\alpha| \le 1).
    • Robust to execution flaws: Adding action noise or timing defects (such as pauses and dropped frames) degrades box-constrained calibration (∣α∣≤1|\alpha| \le 1) by at most 0.02 L1L^1 (0.244-0.265 vs. 0.245 unperturbed), showing it does not overfit jitter.
    • No prior dataset statistics needed: Normalizing actions from the single calibration demo alone matches full-dataset normalization (0.263 vs. 0.265 L1L^1).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.