acceptodds
Under review as a conference paper at ICLR 2027

TacFRS: Plug-in Flow Reversal for High-Frequency Tactile Revision of VLA Models

10 views

Abstract

Vision-language-action (VLA) models offer strong behavioral priors for robotic manipulation, but contact-rich tasks demand rapid feedback that infrequent visual policy updates cannot readily provide. We introduce Tactile Flow Reversal Steering (TacFRS), a lightweight framework that equips pretrained flow-based VLAs with high-frequency tactile action correction while keeping the base policy frozen. TacFRS uses finite-step flow reversal as an action encoding mechanism, mapping predicted action chunks into observation-conditioned structured noise. This representation connects a slow VLA to a fast tactile-conditioned decoder through an asymmetric encoding–decoding architecture. Using this representation, the latest tactile observations, and proprioception, the decoder revises each action between base-policy updates without requiring the VLA to learn a new latent action space. To preserve pretrained behavior while enabling contact-dependent adaptation, we introduce tactile-gated supervision that interpolates between base-policy actions and expert demonstrations according to tactile deviation from a contact-free reference, together with conditional flow matching and endpoint reconstruction objectives. Across four real-world single-arm and bimanual manipulation tasks, TacFRS improves every evaluated task for both backbones, increasing average success from 53.75% to 67.50% for and from 28.75% to 41.25% for SmolVLA. Offline analyses further show that low tactile deviations leave base-policy actions nearly unchanged, whereas high deviations induce larger corrections that reduce ground-truth action error on average.

Primary area:
applications to robotics, autonomy, planning
Keywords:
VLA Models, Tactile Sensing, Flow Reversal, High-Frequency Revision, Visuo-Tactile Manipulation

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_NM2DOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

TacFRS adds high-frequency tactile correction to frozen flow-based VLA policies. It uses flow reversal to convert VLA action chunks into structured noise, then decodes them with a tactile-conditioned revision network. Tactile-gated supervision is used to preserve base-policy behavior when correction is unnecessary. Experiments on four real-robot tasks with π0.5 and SmolVLA show consistent improvements, supported by several ablations.

Critical strengths
  1. The slow-fast design is practical: it adds high-frequency tactile feedback without modifying the pretrained VLA.
  2. The real-robot results are consistent across two backbones, and the ablations support the main design choices.
Critical weaknesses
  1. The structured-noise comparison is not fully convincing. The direct correction baseline is quite weak, and stronger baselines with visual or VLA features would better isolate the benefit of flow reversal.
  2. Evaluation is limited to four tasks, one robot setup, and 20 trials per setting.
Questions for authors
  1. Have you compared against a stronger direct revision baseline using visual or intermediate VLA features?
  2. Are the tactile revision modules trained separately for each task, or can one model handle multiple tasks?
1 comment
AuthorsOpenReview ↗

Response to Reviewer NM2D

Thank you for your constructive feedback and for recognizing the practicality of TacFRS’s slow–fast design and the consistency of our real-robot results across two VLA backbones.

1. Comparison with a direct action refinement baseline that uses visual features

Following your suggestion, we adapted VLA-Touch to use the same fine-tuned π₀.₅ backbone as TacFRS and trained its refinement module with the same number of demonstrations per task (300). The VLA-Touch refinement module conditions on independently extracted DINOv2 RGB features, robot proprioception, and tactile features. Implementation and training details are provided in Appendix D.2, in the VLA-Touch paragraph and Table 12 (p. 27).

We evaluated both methods on all four tasks, with 20 real-robot trials per method per task. The full results are reported in Section 5.2.1 and Figure 5 (p. 7), and the stage-wise success rates are reported in Appendix B.1, Tables 4–7 (pp. 18–19).

MethodPick Two TubesMake BreadInsert TubePress ToasterMean success rate
VLA-Touch (π₀.₅)60%30%65%70%56.25%
TacFRS (π₀.₅)65%35%90%80%67.50%

TacFRS achieved a higher observed success rate on all four tasks, with a mean advantage of 11.25 percentage points.

This comparison measures the overall performance of the two methods. To further examine the role of structured noise obtained through flow reversal, we also evaluated a baseline with the same flow-matching revision network architecture in the structured noise ablation in Section 5.2.4 (pp. 9–10). This baseline initializes the flow from random noise and conditions it on VLA-predicted actions. It achieved success rates of 25% on Insert Tube and 5% on Pick Two Tubes, compared with 90% and 65% for TacFRS, respectively. These results support the use of structured noise obtained through flow reversal in our revision architecture.

2. Can a single tactile revision module serve multiple tasks?

The original experiments used a separately trained tactile revision module for each task. To test whether one module could serve multiple tasks, we jointly trained a shared module on 1,200 demonstrations from all four tasks, with 300 per task, matching the per-task count in the original experiments. Each task retained its own frozen π₀.₅ policy and used the same tactile revision module weights. The shared module uses a common 20-dimensional bimanual action representation and masks the action dimensions and tactile inputs of the inactive arm on single-arm tasks.

We evaluated the same jointly trained checkpoint on all four tasks, with 20 real-robot trials per task. The training setup and results are described in Section 5.2.2 (pp. 7–8), with the full comparison in Table 2 (p. 8).

MethodPick Two TubesMake BreadInsert TubePress ToasterMean success rate
π₀.₅50%20%75%70%53.75%
π₀.₅ + TacFRS (task-specific)65%35%90%80%67.50%
π₀.₅ + TacFRS (jointly trained on four tasks)60%25%65%70%55.00%

The same tactile revision module could be used across all four evaluated tasks. Its mean success rate was 55.00%, 1.25 percentage points above the frozen π₀.₅ policies and below the 67.50% achieved by task-specific training. We discuss this result in Section 5.2.2 of the revised manuscript.

3. Number of real-robot trials and evaluation across robot embodiments

We agree that the evaluation scope is limited. Our real-robot experiments cover four tasks on a single robot platform, with 20 trials per method per task (Section 5.1, pp. 6–7). Given the time available during the rebuttal period, we focused our additional experiments on the VLA-Touch comparison and the evaluation of the shared tactile revision module. We were unable to increase the number of trials per setting or run experiments on additional robot embodiments within that period. Our conclusions are based on the platform and tasks evaluated here. Transfer across robot embodiments remains to be tested, as discussed in the Limitations paragraph in Section 6 (p. 10).

Reviewer_UJpYOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TacFRS, a plug-in framework for tactile-guided action revision of pretrained flow-based VLA policies. It applies flow reversal to map VLA-generated actions into structured noise, which is then used by a tactile-conditioned flow decoder to generate revised actions at a higher frequency. Experiments on four real-world manipulation tasks with π0.5 and SmolVLA demonstrate consistent improvements in task success rates over the corresponding base policies.

Critical strengths
  1. Practical and modular tactile revision framework. TacFRS introduces a plug-in tactile revision mechanism for pretrained VLA policies, enabling contact-aware action refinement without modifying the base policy. The modular design provides a practical approach to incorporating tactile feedback into robotic manipulation.

  2. Consistent improvements in real-world manipulation. The framework demonstrates consistent improvements in task success across multiple real-world manipulation tasks and VLA backbones, supporting its practical effectiveness.

Critical weaknesses
  1. Insufficient isolation of the structured noise contribution. The direct correction baseline regresses corrected actions from VLA predictions and tactile embeddings, but achieves 0% success across all four tasks, substantially underperforming the original VLA. This raises concerns about whether it constitutes a sufficiently competitive action-refinement baseline. Since the comparison also changes the action representation and generation mechanism, the reported performance gap does not convincingly isolate the benefit of structured noise.

  2. Insufficient comparison with closely related generative refinement methods. Prior work such as VLA-Touch has explored tactile-conditioned generative refinement of VLA actions without flow reversal. The current comparisons do not sufficiently establish whether the proposed flow-reversal interface offers advantages over directly refining VLA-generated actions with a generative model.

Questions for authors
  1. Why does the direct correction baseline achieve 0% success across all four tasks? Does it preserve the original VLA actions through residual or skip connections, and could the authors evaluate a stronger residual correction baseline?

  2. Could the authors compare structured noise with VLA-generated actions using the same tactile-conditioned flow decoder, training objectives, and inference budget to isolate the contribution of flow reversal?

  3. Could the authors provide a controlled comparison with VLA-Touch-style generative action refinement to clarify the advantages of the proposed flow-reversal interface?

2 comments

We thank the reviewer for the careful reading and constructive feedback. Your comments have helped us strengthen the baseline comparisons and clarify the contribution of the flow-reversal interface. We provide point-by-point responses below.

Critical Weaknesses:

1. Insufficient isolation of the structured noise contribution. The direct correction baseline regresses corrected actions from VLA predictions and tactile embeddings, but achieves 0% success across all four tasks, substantially underperforming the original VLA. This raises concerns about whether it constitutes a sufficiently competitive action-refinement baseline. Since the comparison also changes the action representation and generation mechanism, the reported performance gap does not convincingly isolate the benefit of structured noise.

Answer:

Thank you for pointing this out. The original direct correction baseline changed both the action representation and the generation mechanism: it fed VLA-predicted actions and tactile features into a Transformer to directly predict revised actions, rather than a flow-matching velocity field. Its 0% success rate therefore does not, by itself, isolate the contribution of structured noise obtained through flow reversal.

We added a flow-matching baseline that initializes the flow from random noise and conditions the velocity field on VLA-predicted action chunks, together with tactile observations and proprioceptive states. It uses the same π0.5\pi_{0.5} backbone, decoder architecture, tactile and proprioceptive inputs, training data, tactile-adaptive supervision, flow-matching and endpoint reconstruction losses, number of decoding integration steps, and control frequency as TacFRS. The comparison concerns how VLA predictions are incorporated into the TR network: the baseline uses them as explicit conditioning inputs, whereas TacFRS converts them into structured noise through observation-conditioned flow reversal and uses that noise to initialize the flow.

On Pick Two Tubes and Insert Tube, the new baseline achieves success rates of 5% and 25%, respectively, compared with 65% and 90% for TacFRS. By holding the flow-matching formulation, training settings, and decoding budget fixed, this comparison evaluates the flow-reversal interface while removing the confound introduced by changing the generation mechanism in the original experiment. Flow-matching refinement from random noise conditioned on VLA-predicted actions substantially underperforms TacFRS on both tasks, further supporting the importance of the flow-reversal interface for achieving the reported performance.

2. Insufficient comparison with closely related generative refinement methods. Prior work such as VLA-Touch has explored tactile-conditioned generative refinement of VLA actions without flow reversal. The current comparisons do not sufficiently establish whether the proposed flow-reversal interface offers advantages over directly refining VLA-generated actions with a generative model.

Answer:

Thank you for highlighting this missing comparison. VLA-Touch is closely related to our work, and comparisons with other tactile-fusion methods alone do not establish the advantage of TacFRS over direct generative action refinement. We therefore added an adaptation of VLA-Touch with a π0.5\pi_{0.5} backbone. Its tactile-conditioned action-refinement module was trained on Insert Tube and Pick Two Tubes using the same training data and evaluation settings as TacFRS. The results are reported in Figure 5 and Appendix B.1 of the revised manuscript.

On Insert Tube, we observed that the DeepSeek planner sometimes misidentified the task stage under the limited wrist-camera view, causing premature gripper closure. These failures reflect the influence of high-level planning in our VLA-Touch implementation and cannot be directly attributed to generative refinement. TacFRS uses continuously updated tactile feedback to adjust actions after contact.

Pick Two Tubes requires sequential bimanual grasping and extraction under changes in the spatial layout. VLA-Touch improves the success rate from 50% for the base π0.5\pi_{0.5} policy to 60%, demonstrating the value of direct tactile refinement; TacFRS further increases it to 65%. These results support its effectiveness under the tested conditions, without establishing broader generalization.

This comparison adds a direct generative refinement baseline and shows higher success rates for TacFRS on both tasks, strengthening the evidence for the overall framework. To examine the contribution of flow reversal, we also conducted the controlled comparison described above, initializing the flow from random noise and conditioning the velocity field on VLA-predicted action chunks while keeping the decoder architecture, training objectives, and decoding budget fixed. TacFRS also achieves higher success rates in this comparison, further supporting the design choice of using the flow-reversal interface.

Questions For Authors:

1. Question: Why does the direct correction baseline achieve 0% success across all four tasks? Does it preserve the original VLA actions through residual or skip connections, and could the authors evaluate a stronger residual correction baseline?

Answer:

Thank you for the suggestion. The original direct correction baseline fed VLA-predicted actions and tactile features into a Transformer to directly predict revised actions, rather than a flow-matching velocity field. Although the original VLA actions were provided as input, the output could still deviate from the intended task execution, and the baseline achieved 0% task completion across all four tasks. For example, on Press Toaster, it skipped target localization and pressed downward along a steeper trajectory at a higher speed. One possible explanation is that VLA-predicted actions and tactile features provide insufficient scene context to distinguish the corrections needed in different situations.

To provide a stronger action-refinement comparison, we additionally evaluated VLA-Touch with a π0.5\pi_{0.5} backbone. It achieves success rates of 65% and 60% on Insert Tube and Pick Two Tubes, respectively, compared with 90% and 65% for TacFRS. The advantage of TacFRS therefore also holds against a generative refinement baseline that successfully completes these tasks.

2. Question: Could the authors compare structured noise with VLA-generated actions using the same tactile-conditioned flow decoder, training objectives, and inference budget to isolate the contribution of flow reversal?

Answer:

Thank you for the suggestion. The original direct correction baseline fed VLA-predicted actions and tactile features into a Transformer to directly predict revised actions, rather than a flow-matching velocity field. Although the original VLA actions were provided as input, the output could still deviate from the intended task execution, and the baseline achieved 0% task completion across all four tasks. For example, on Press Toaster, it skipped target localization and pressed downward along a steeper trajectory at a higher speed. One possible explanation is that VLA-predicted actions and tactile features provide insufficient scene context to distinguish the corrections needed in different situations. TacFRS obtains structured noise containing action and visual information through observation-conditioned flow reversal, providing scene context for the tactile revision network to make context-appropriate adjustments.

We have revised this baseline to use the same flow-matching TR architecture and training objectives as TacFRS. Specifically, the revised baseline initializes the flow from random noise and conditions the velocity field on the VLA-predicted action chunk, together with tactile observations and proprioceptive states. This keeps the flow-matching policy formulation unchanged and compares explicit conditioning on VLA actions with the flow-reversal interface used by TacFRS.

To further strengthen the action-refinement comparison, we additionally evaluated VLA-Touch with a π0.5\pi_{0.5} backbone. It achieves success rates of 65% and 60% on Insert Tube and Pick Two Tubes, respectively, compared with 90% and 65% for TacFRS. The advantage of TacFRS therefore also holds against a generative refinement baseline that successfully completes these tasks.

3. Question: Could the authors provide a controlled comparison with VLA-Touch-style generative action refinement to clarify the advantages of the proposed flow-reversal interface?

Answer:

Thank you for the suggestion. We added VLA-Touch with a π0.5\pi_{0.5} backbone and evaluated it using the same training data and test settings as TacFRS. On Insert Tube and Pick Two Tubes, VLA-Touch achieves success rates of 65% and 60%, respectively, while TacFRS achieves 90% and 65%, corresponding to improvements of 25 and 5 percentage points.

VLA-Touch uses predicted actions as the starting point for generative refinement. TacFRS instead encodes actions into structured noise through finite-step flow reversal conditioned on the original observations, then decodes this representation at high frequency using the latest tactile feedback and proprioceptive state. These results provide additional evidence for the overall performance of TacFRS relative to direct generative refinement. Since the VLA-Touch comparison also involves implementation differences, including high-level planning, we additionally conducted the controlled flow-matching comparison described above, initializing the flow from random noise and conditioning the velocity field on VLA-predicted action chunks while holding the decoder architecture, training objectives, and decoding budget fixed to further examine the contribution of flow reversal.

Reviewer_G17YOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the problem of adding fast tactile feedback to a pretrained flow-based VLA, which typically executes action chunks open-loop and is updated at a low frequency, without retraining the VLA itself.

The proposed method, TacFRS, keeps the VLA frozen and uses it as a slow encoder. At each VLA update, the generated action chunk is integrated backwards through the same observation-conditioned flow to obtain what the authors call structured noise, and a separate lightweight tactile revision network, itself a flow-matching model conditioned on tactile history and proprioception, repeatedly decodes this structured noise at 30 Hz so that each executed action reflects the latest contact information. The revision network is trained on an additional set of demonstrations with a target that interpolates between the frozen VLA action and the demonstrated action, where the interpolation weight is a gate computed from how far the current tactile signal deviates from a contact-free reference.

On four real-robot tasks, TacFRS raises the average success rate of π0.5 from 53.75% to 67.5% and that of SmolVLA from 28.75% to 41.25%, and with π0.5 it also outperforms the visuo-tactile baselines DECO and RDP.

Critical strengths

1. Practicality of the proposed design. TacFRS adds tactile feedback to a pretrained flow-based VLA while keeping the VLA frozen, so that the behavioral priors learned from large-scale vision-only data are preserved and only a lightweight revision network has to be trained on a few hundred additional demonstrations per task. This is consequential for my assessment because tactile demonstrations are much harder to collect at scale than vision-only ones, and a plug-in approach that reuses existing VLAs addresses a real practical need in contact-rich manipulation.

2. The method is evaluated on real hardware with two different backbones and improves both of them on all four tasks, including two bimanual tasks. The appendix also reports stage-wise completion rates for all six methods, which gives a more detailed picture of where the improvements occur than final success rates alone. This is consequential because the consistent direction of the results across backbones and tasks suggests that the approach is not tied to a particular policy.

Critical weaknesses
  1. The first critical weakness is that the contribution of the tactile signal itself is not isolated. The revision network is trained with supervision from demonstrated actions on 300 additional demonstrations per task and re-decodes the action at every control step with fresh proprioception, and in the reported validation analysis 2,778 of the 3,219 samples fall into the high-gate group, where the training target is dominated by the demonstrated action. The improvements could therefore come, at least in part, from an additional stage of supervised refinement or from closed-loop re-decoding rather than from touch.

However, none of the reported ablations removes the tactile input, and the paper compares neither with the same revision network trained without touch nor with non-tactile flow reversal steering as in FRS (Tang et al., 2026), on which the method is explicitly built. This is consequential because the benefit of touch is the central claim of the paper, and the current results remain consistent with the simpler explanation that any additional revision stage improves the frozen VLA.

  1. The second critical weakness is that the necessity of the flow-reversal interface, which is the main methodological novelty, is not convincingly demonstrated. The only alternative interface is a direct correction model that differs from TacFRS in several respects at once and achieves 0% success on all four tasks, which, given that this model has access to the VLA's own actions, suggests an insufficiently tuned baseline rather than evidence in favour of the proposed interface.
Questions for authors

Clarification requests

  1. The structured noise is fixed for the duration of each chunk of about 1.7 seconds, and the revision network receives tactile and proprioceptive input but no visual input. Are the corrections therefore effectively confined to a neighborhood of the original VLA plan, and how does the system behave when contact calls for a change of plan within a chunk, for instance a re-grasp after the object slips?

  2. Is a separate revision network trained for each task, and if so, would a single revision network trained on all tasks retain the improvements? This matters for the claim that TacFRS is a plug-in module for existing VLAs.

  3. Since the reported frequencies are configured targets rather than measured throughput, what are the measured on-robot latencies of the reverse integration and of each call of the revision network, and does the 30 Hz revision actually run at that rate?

Analyses that could change my assessment

  1. How does TacFRS perform when the revision network is trained and deployed with the tactile input replaced by the contact-free reference, and how does a non-tactile FRS-style steering of the same frozen VLA perform on these tasks? Relatedly, which components of TacFRS are new relative to FRS?

  2. How does a revision network perform when it keeps the same flow-matching decoder and training objective but is conditioned on the VLA action chunk instead of being initialized from the structured noise, so that only the interface differs from TacFRS?

  3. The tactile gate measures deviation from a contact-free reference rather than whether a correction is actually needed, as the authors acknowledge. How often does the gate activate during contact that requires no correction, such as stable grasping, and does the revision network modify the actions substantially in those cases?

4 comments

We thank the reviewer for the careful reading of our manuscript and the detailed, constructive feedback. Your comments have helped us clarify the method’s scope and limitations and strengthen the experimental evaluation. We have revised the manuscript accordingly and provide point-by-point responses below. The corresponding revisions are highlighted in red in the revised manuscript.

1. Question: The structured noise is fixed for the duration of each chunk of about 1.7 seconds, and the revision network receives tactile and proprioceptive input but no visual input. Are the corrections therefore effectively confined to a neighborhood of the original VLA plan and how does the system behave when contact calls for a change of plan within a chunk, for instance a re-grasp after the object slips?

Answer:

To address this question, Fig. 8(a) shows the VLA-generated actions (dashed gray), revised actions (solid blue), and measured robot pose (black) during an Insert Tube rollout. Each short gray or blue segment shows the sequence of commands in one action chunk. Since each action specifies a pose increment relative to the previous time step, we reconstruct the commanded trajectories by composing these increments from the reference pose at the start of each chunk. The two command traces closely overlap for most of the rollout, with visible deviations when the tactile revision network adjusts the commands. We have added this analysis to Section 5.2.3, and the revisions are shown in red in the manuscript.

When contact calls for a change of plan within a chunk, TacFRS can revise the VLA-generated actions, but it inherits the VLA’s replanning frequency. A re-grasp that requires visually relocalizing the object must therefore wait until the next action chunk update. TacFRS can still make local adjustments to the planned trajectory. For example, in the Make Bread task, TacFRS adjusted the robot’s grip so that it held the container more firmly (Appendix B.3, Fig. 11).

2. Question: Is a separate revision network trained for each task, and if so, would a single revision network trained on all tasks retain the improvements? This matters for the claim that TacFRS is a plug-in module for existing VLAs.

Answer:

To answer this question, we jointly trained a single tactile revision module on 1,200 demonstrations from all four tasks, using 300 demonstrations per task to match the original per-task training budget. Each task retained its own frozen π0.5\pi_{0.5} policy and used the same shared revision module. The module uses a common 20-dimensional bimanual action representation and masks the inactive arm’s action dimensions and tactile inputs on single-arm tasks.

We evaluated the same checkpoint on Insert Tube and Pick Two Tubes in real-robot trials. On Insert Tube, it achieved 65% (13/20) success, compared with 75% for π0.5\pi_{0.5} alone and 90% for task-specific TacFRS. On Pick Two Tubes, it achieved 60% (12/20), compared with 50% for π0.5\pi_{0.5} alone and 65% for task-specific TacFRS. For the remaining tasks, Table 2 reports 25% (5/20) success on Make Bread and 70% (14/20) on Press Toaster for the jointly trained module, compared with 20% and 70% for π0.5\pi_{0.5} alone and 35% and 80% for task-specific TacFRS, respectively. Across all four tasks, the jointly trained module achieved an average success rate of 55.00%, compared with 53.75% for π0.5\pi_{0.5} alone and 67.50% for task-specific TacFRS.

This result suggests that a single revision network could serve as a plug-in module for general-purpose action correction across tasks, although the jointly trained module did not retain the performance gains achieved through task-specific training. We have added these results to Section 5.2.2 (Table 2), and the revisions are shown in red in the manuscript.

3. Question: Since the reported frequencies are configured targets rather than measured throughput, what are the measured on-robot latencies of the reverse integration and of each call of the revision network, and does the 30 Hz revision actually run at that rate?

Answer:

We clarify that 30 Hz refers to the target intra-chunk tactile-revision update rate. Measurements were performed on an NVIDIA RTX 5080 using high-resolution monotonic wall-clock timers and JAX synchronization. Each test condition included 10 chunks of 50 actions, providing 10 reverse-integration measurements and 500 revision measurements. Under continuous execution, reverse integration required 196.29±1.46196.29 \pm 1.46 ms including preparation, with 152.49±1.05152.49 \pm 1.05 ms attributable to integration itself. The structured noise is computed once and cached for all 50 actions, so this initialization cost is not repeated during individual revisions. The tactile revision network contains 7.52M trainable parameters, and its full revision pipeline requires 15.74±26.2515.74 \pm 26.25 ms per update. With revisions scheduled at 30 Hz, the measured average update rate within each action chunk was 29.76 Hz, based on 500 updates. We have added these measurements to Appendix D.5. The revisions are shown in red in the manuscript.

4. Critical Weaknesses: The first critical weakness is that the contribution of the tactile signal itself is not isolated. The revision network is trained with supervision from demonstrated actions on 300 additional demonstrations per task and re-decodes the action at every control step with fresh proprioception, and in the reported validation analysis 2,778 of the 3,219 samples fall into the high-gate group, where the training target is dominated by the demonstrated action. The improvements could therefore come, at least in part, from an additional stage of supervised refinement or from closed-loop re-decoding rather than from touch.

However, none of the reported ablations removes the tactile input, and the paper compares neither with the same revision network trained without touch nor with non-tactile flow reversal steering as in FRS (Tang et al., 2026), on which the method is explicitly built. This is consequential because the benefit of touch is the central claim of the paper, and the current results remain consistent with the simpler explanation that any additional revision stage improves the frozen VLA.

Question: How does TacFRS perform when the revision network is trained and deployed with the tactile input replaced by the contact-free reference, and how does a non-tactile FRS-style steering of the same frozen VLA perform on these tasks? Relatedly, which components of TacFRS are new relative to FRS?

Answer:

Thank you for this suggestion. To verify the importance of tactile information for action correction and demonstrate that closed-loop re-decoding alone is insufficient to improve actions, we added an ablation in which the revision network retained the VLA-derived structured-noise input and was conditioned only on proprioceptive states. Furthermore, we retain the original loss formulation in this baseline, since a network supervised solely by VLA-predicted actions would effectively become the action-reconstruction decoder described in the appendix.

This baseline therefore satisfies both requirements: removing tactile input and performing non-tactile FRS-style action correction. We evaluated this variant on Tasks 1 and 3. The network learned almost no effective actions. In most trials, the robot arm remained stationary. In a few cases, it moved forward briefly and then closed the gripper. The success rate was 0% on both tasks. We have added these results to Section 5.2.4 (Tactile Information Ablation), and the revisions are shown in red in the manuscript.

Regarding the differences from FRS (Tang et al., 2026), its zero-shot steering pipeline uses a VLM or human reasoner to infer a coarse motion, which a predefined procedure converts into a reference action chunk. The frozen VLA then approximately inverts this chunk through numerical reverse integration and uses the resulting noise to initialize forward action generation. With finite-step integration, the generated actions approximately follow the reference while reflecting the VLA’s learned behavioral prior. A key reason why FRS works without additional training is therefore that the network decoding the structured noise already possesses a strong behavioral prior, allowing even very coarse VLM-generated reference actions to steer it toward improved actions.

We build on the idea of using flow-reversed noise to steer action generation. In TacFRS, we invert the VLA’s own predicted action chunks and decode the resulting structured noise using a smaller tactile revision network. Our revision network must therefore be trained separately with informative inputs. Our reconstruction experiment in Appendix C.3 provides empirical evidence that this representation retains sufficient information for a separately trained decoder to approximately reconstruct the source actions without observation conditioning. To balance this reconstruction ability with tactile correction, we introduce tactile-gated supervision, which blends VLA and expert actions according to tactile deviation from a contact-free reference. We also introduce a slow–fast policy, in which the revision network repeatedly decodes the structured noise using fresh tactile and proprioceptive feedback between VLA updates.

5. Critical Weaknesses: The second critical weakness is that the necessity of the flow-reversal interface, which is the main methodological novelty, is not convincingly demonstrated. The only alternative interface is a direct correction model that differs from TacFRS in several respects at once and achieves 0% success on all four tasks, which, given that this model has access to the VLA's own actions, suggests an insufficiently tuned baseline rather than evidence in favour of the proposed interface.

Question: How does a revision network perform when it keeps the same flow-matching decoder and training objective but is conditioned on the VLA action chunk instead of being initialized from the structured noise, so that only the interface differs from TacFRS?

Answer:

We agree that our original direct correction baseline did not isolate the effect of the flow-reversal interface. The baseline fed VLA actions into a network that directly predicted revised actions, changing both the flow-reversal interface and the policy formulation. In the revised baseline, we retain the same flow-matching TR architecture but initialize the flow from random noise and condition the velocity field on the VLA-predicted action chunk, together with tactile observations and proprioceptive states. This keeps the flow-matching policy formulation unchanged and compares explicit conditioning on VLA actions with the flow-reversal interface used by TacFRS. Using π0.5\pi_{0.5} as the base VLA model, we evaluated both methods on Tasks 1 and 3. On Insert Tube, the revised baseline achieved a success rate of 25%, compared with 90% for TacFRS. On Pick Two Tubes, it achieved 5%, compared with 65% for TacFRS. These results support our conclusion that the flow-reversal interface is important for achieving TacFRS’s performance on these tasks. We have added these results to Section 5.2.4 (Structured Noise Ablation). The revisions are shown in red in the manuscript.

6. Question: The tactile gate measures deviation from a contact-free reference rather than whether a correction is actually needed, as the authors acknowledge. How often does the gate activate during contact that requires no correction, such as stable grasping, and does the revision network modify the actions substantially in those cases?

Answer:

To address this question, we plot the pose deviation and the cosine similarity between the tactile token and its contact-free reference on a shared timeline for a Two-Tube Picking rollout (Fig. 8(b)). We define the pose deviation between the VLA-generated and revised actions as

\left( \frac{\lVert p_{\mathrm{revised}} - p_{\mathrm{VLA}}\rVert_2}{10\\,\mathrm{mm}} \right)^2 + \left( \frac{\operatorname{angle}(R_{\mathrm{VLA}}^{\top} R_{\mathrm{revised}})}{10^{\circ}} \right)^2 }$$ This dimensionless metric normalizes the translation and rotation differences by $10\\,\mathrm{mm}$ and $10^{\circ}$, respectively. Each short blue curve shows the pose deviation over one action chunk. When the left gripper grasped the tube and pulled it from the rack, tactile similarity dropped sharply and TacFRS made a large pose correction. At the correction peak, $D$ reached 6.5454, with translation and rotation deviations of 63.401 mm and 16.266°, respectively. As the robot retracted toward its initial pose while still holding the tube, tactile similarity remained low and the pose corrections were small. During this stable-grasping phase, the mean $D$ over the final six action chunks was 0.3145, with mean translation and rotation deviations of 2.429 mm and 1.608°, respectively. This suggests that TacFRS uses tactile feedback in the context of the VLA’s intended motion, so similar levels of tactile deviation can lead to different amounts of correction. We have added these results to Section 5.2.3, and the revisions are shown in red in the manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.