- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 1
- Soundness
- 1
Read review
Paper summary
This paper mainly studies how multiple pretrained vision-language-action models can coordinate during multi-arm manipulation. The paper proposes VLA-Fabric, which assigns one complete VLA to each arm and exchanges information between the models through shared context, intermediate features, and action-stage interaction. The authors compare different interaction designs, implement the method in SingleVLA and π₀.₅, and evaluate it on bimanual and larger-team tasks in simulation and on physical robots.
Critical strengths
- The paper compares different interaction locations, fusion operators, and fine-tuning strategies, which helps in understanding how these choices affect coordination.
- The experiments cover two VLA architectures and include three- and four-arm tasks in both simulation and the real world. The handover and shared-object manipulation results provide useful evidence that the method can support coordinated execution.
Critical weaknesses
- The Full results in Tables 1A and 1B are different. For example, SingleVLA-Fabric achieves 80.0% versus 75.0% on Box and 70.0% versus 78.0% on Shoes. The Box result for π₀.₅-Fabric also changes from 80.5% to 79.0%. Appendix A.5 mentions a separate evaluation batch for the SingleVLA ablations, but does not clearly explain what changed between the two evaluations. Since Table 1B only disables interaction during inference without retraining, it is unclear to me whether the same trained checkpoint was used in both parts and what caused the different results.
- Tables 2 and 3 only compare Full with Independent, both of which already use a separate VLA for each arm. These comparisons show the benefit of information exchange, but it remains unclear whether a single VLA receiving all arms’ observations and states could achieve similar results by jointly predicting their actions. Moreover, the four-arm system uses four complete VLA copies and additional interaction modules. The paper does not sufficiently report their parameter count, computational cost, memory requirements, or communication overhead, so it is difficult to assess whether the performance justifies the additional resources. The bimanual comparisons in Table 1 do not answer this question for the larger-team and real-world tasks.
- The practical benefit of modularity remains unclear to me. Sections 5 and 7 emphasize keeping local action dimensions unchanged and re-instantiating coordination across VLA architectures. However, Appendix C.1 states that each task and team configuration is fine-tuned separately, and the SingleVLA and π₀.₅ versions use architecture-specific interaction implementations. These results support reuse of the design principle, but do not establish how much adaptation effort is saved when changing the team configuration or VLA architecture. Could the authors clarify what can be reused and what needs to be redesigned or fine-tuned?
Questions for authors
- Could the authors explain why the Full results differ between Tables 1A and 1B? Please clarify whether the same checkpoints, test conditions, and inference settings were used. If different checkpoints were used, I would appreciate an explanation of why this was needed for the fixed-weight ablation study.
- I would suggest adding a joint VLA baseline that receives all arms’ observations and states and jointly predicts their actions. The comparison should use the same observation sources, demonstrations, and evaluation conditions, with comparable pretraining and training budgets. Reporting the total parameter count, training compute, peak memory usage, and inference latency would also help in assessing the performance relative to the computational cost.
4 comments
Response Part 2/3 | W3 | Modularity and reuse
Weaknesses (continued)
W3. Thank you for asking what carries over. We chose π0.5 after studying SingleVLA because their internal computation differs substantially. As Figure 3 shows, SingleVLA first processes observations in its VLM, then fuses the readouts before separate local action decoders run. In π0.5, Common acts on the multimodal prefix, peer Private K/V condition the Action Expert, and action features are exchanged as the Action Expert computes each flow step. The transfer shows that the same three functions can be placed at different computational boundaries.
To apply the design to a new backbone, we identify the features that can serve as shared context, peer-specific information, and action-side intent. The placement, operators, and learned weights are then tailored to that backbone. Across team sizes, each arm keeps its observation and action interfaces, while the peer set and some fusion components change and the team is fine-tuned. At fixed team size, Shared π0.5-Fabric reuses one checkpoint per role across four bimanual tasks. This gives us a concrete way to extend the team by keeping the existing arms’ input and action interfaces and fine-tuning the team to coordinate with the additional arm. Our experiments establish the adapted teams’ task performance. Thank you for raising the question of how much of a trained team can be reused when adding an arm. It gives us a concrete direction for studying how to preserve the existing arms’ skills as the team expands.
Response Part 3/3 | Q1-Q2 | Checkpoints, joint π0.5, and resource measurements
Questions
Thank you for specifying which evaluation settings and resource measurements would make the comparisons more useful.
Q1. For SingleVLA Box and Shoes, the A and B entries use different Full checkpoints and condition sets. On Box, the Full run in A achieved 160/200. The Full run in B achieved 150/200. On Shoes, the Full run in A achieved 140/200. The Full run in B achieved 156/200. For each removal in B, we hold the checkpoint, test conditions, and other inference settings fixed relative to its own Full reference.
For SingleVLA Mic, both entries point to the same archived checkpoint path and use different condition sets. The Full run in A achieved 168/200. The Full run in B achieved 163/200. For π0.5 Box, we verified that the original runs used the same recorded checkpoint, seed, conditions, and Full inference settings. The run in A achieved 161/200 (80.5%). The run in B achieved 158/200 (79.0%).
To check where variability can arise, we held the weights, input, and actual flow noise fixed. Two fresh inference processes still differed slightly in their outputs, with maximum normalized action differences of 0.00277 for the left arm and 0.00168 for the right. In a check with compiler autotuning disabled, two fresh processes produced identical outputs for that input. Rendering the same simulator state without advancing it also produced occasional RGB differences of one intensity level (1/255). These diagnostics point to numerical differences between inference processes and small rendering changes as possible contributors to the Box score differences.
We reloaded the original π0.5 Box checkpoint and evaluated it twice on the same 200 conditions. The two repeat evaluations achieved 77.5% and 79.0%, respectively. Across repeats, the weights, initial positions and velocities, proprioceptive states, and first planning-step sampling keys matched. The repeats show that success counts can fluctuate slightly under the same recorded setup. We retain both original table entries.
Q2. The joint π0.5 uses the shared scene image, all wrist images, and all arm states from the same observation sources as Fabric. We adapted image preprocessing to retain those views and used the native 32-dimensional action head for joint prediction.
| Task | Fabric Full | Independent | Joint π0.5 |
|---|---|---|---|
| Frame3 | 81.5% | 76.5% | 74.0% |
| Frame4 | 97.0% | 39.0% | 73.5% |
Full outperformed Independent on both tasks. The joint results in W2 are close to Full on Frame3, with a larger gap on Frame4.
For the joint comparisons we used the same demonstration source, base π0.5 checkpoint, and update budget as Fabric Full within each task.
The four-arm Frame4 Full run on four A800 GPUs averaged 2.09 seconds per optimizer update near the end of training and reached a peak JAX allocator footprint of 31.60 GiB per GPU. Those measurements characterize training. For inference, we can share the communication and planning measurements from the separate study below.
Communication optimization for VLA-Fabric is the focus of a separate study of ours. In four-agent Eagle2/SingleVLA Take Photo, the RAW and optimized runs used the same checkpoint and 50 initial conditions. Team-wide logical payload fell from 310.785 to 34.535 MiB per planning round (88.89%), and mean planning latency fell from 875.5 to 322.4 ms per round. These measurements are for that Eagle2 setup. We appreciate your attention to this practical question and look forward to sharing the broader communication results in that study.
Comment by Reviewer yQP6
Thank you for the response and additional experiments. The explanation of Table 1 and the new joint π₀.₅ comparisons address several of my concerns. However, the training and inference costs relative to the joint policy remain insufficiently evaluated. It is also still unclear how much adaptation effort the proposed modularity saves when changing the team configuration, since no corresponding reuse experiment is provided. I therefore maintain my original rating.
Response Part 1/3 | W1-W2 | Evaluation and deployment costs
Thank you for examining the experiments. Please allow us to briefly restate the paper’s main contribution.
VLA-Fabric explores a way to organize complete VLAs for multi-arm learning through explicit exchange of their internal representations. We establish the interaction architecture in SingleVLA, transfer it to the different computation of π0.5, and apply it to teams of two to four arms in simulation. The performance comparisons evaluate this initial design under familiar VLA fine-tuning objectives. The architecture gives us a basis to combine improvements in local policies, exchanged representations, and communication efficiency, with room to improve coordination and task performance.
During the discussion period, we consolidated the reviews and added experiments to address them. We completed four larger-team joint π0.5 evaluations, two CHORUS configurations, bimanual π0.5 Independent and joint controls, and wrist-only vision evaluations. We draw on these experiments below to address the specific points in your review. We will use the ten-page main text and appendix in the discussion revision to explain the method, protocols, and results. Thank you for guiding these additions.
Weaknesses
W1. Thank you for asking us to clarify Appendix A.5. During development we trained multiple Full checkpoints with the same interaction design. For SingleVLA Box and Shoes, Table 1A reports the capability runs. In Table 1B, we selected another Full checkpoint for each task to check whether reliance on the exchange paths also appeared in a different trained instance of the same design. Those diagnostics used a separate batch of test conditions. Q1 details the checkpoint and condition records, including the π0.5 Box repeats.
Within Table 1B, each removal uses the corresponding Full checkpoint and test conditions. On Handover Mic, Full π0.5-Fabric achieves 97.5%. With Private K/V removed, the same policy achieves 11.0%. Independent trains a separate policy per arm without interaction modules. The Frame4 results in Q2 illustrate how Full and Independent compare on a task that requires the arms to move a shared object together.
W2. Thank you for asking whether one VLA could achieve comparable performance with more arms. Table 1 compares bimanual performance. To address the larger-team question, we added a joint π0.5 on each larger-team simulation task. One model sees all arms' observations and states and outputs their actions together.
Following suggestions from other reviewers, we also evaluated both CHORUS configurations and Role-mixed Fabric. We include these results here for your reference.
The results show that a joint model can learn some of these tasks, with larger gaps on tasks such as Frame4.
The native action head also constrains which teams a single model can control. Four 7-dimensional arm actions use 28 of π0.5's 32 output dimensions, while a fifth would require 35 with the same interface. SingleVLA's 10-dimensional head was designed for one arm, so it would need a different output interface even for joint bimanual control. Fabric adds one complete VLA per arm and keeps the existing local action heads. The four-arm Fabric configuration has about 14.067B total parameters, or 3.517B per device with one VLA per arm, versus 3.403B for joint π0.5. Q2 gives the training measurements and a communication result from our separate optimization study.