acceptodds
Under review as a conference paper at ICLR 2027

SPECTRA: Gated Spectral Policies for Force-Aware Manipulation

12 views

Abstract

Robot manipulation requires smooth movements as well as precise adjustments. Force and tactile feedback can help a robot understand its interaction with an object, especially when visual observations alone are insufficient. However, this feedback is not equally useful throughout a task. Similarly, predicting finer motion details can help the robot make necessary corrections, but can also reproduce noise in demonstrations. An effective policy should therefore learn when force feedback and finer motion are useful. We introduce SPECTRA, a force-aware spectral diffusion policy that makes this choice during execution. It combines a coarse policy for smooth, low-frequency motion with a fine policy that uses additional force/tactile feedback and can predict the full range of motion frequencies. A learned gate uses recent force history to select the policy that generates the next action chunk. Experiments on simulated and real-robot manipulation tasks show that SPECTRA improves real-robot task success by up to 20 percentage points over the strongest evaluated baseline on the corresponding task. Comparisons with individual policies further show that the benefit of force feedback depends on action bandwidth, supporting its selective use with fine motion prediction.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Robot Learning, Robot Manipulation, Behavior Cloning

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_5AYCOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

SPECTRA is a force-aware diffusion policy that trains two independent experts: a force-blind coarse policy restricted to the lowest k=3 DCT coefficients, and a force-conditioned fine policy that predicts the full spectrum. Per-coefficient standardization (“EqualSNR”) is applied before diffusion. A small MLP gate sees only the recent force history and picks one expert per action chunk. The gate is trained on which frozen expert has lower offline prediction error on held-out demos, so no contact labels are needed. Experiments use six simulated tasks (ManiSkill ports in MuJoCo, plus MimicGen) and five real OpenArm tasks.

Critical strengths
  • Clear, well-motivated idea. Adding force is useful for contact-rich manipulation, and its usefulness depends on whether the policy can express fast corrections. The idea to combine force input with high-frequency actions is a sensible framing. The method is also a simple, modular design without much overhead.
  • SPECTRA beats both of its own experts. Comparing Fig. 3 and Table 1, it beats both the coarse-only and fine-only policies on every simulated task shown.
  • Informative ablations. The comparisons cross backbone (DP/FM), representation (action/DCT), and force strategy (none/always/gated), with full real-robot tables in the appendix.
Critical weaknesses
  • SPECTRA shows limited improvement over success rates over most real world tasks as shown in Figure 5.
  • The hypothesis H4 is under-evaluated. Section 4.1.1 promises fixed or contact-triggered routing comparisons with the same experts, but no results appear. The “generic gating” baseline uses different policies, so it is not a controlled comparison.
  • Lack of comparisons with existing force-aware policies. There's no comparison with prior works like Reactive Diffusion Policy[1], ACP[2], making it hard to justify the effectiveness of the paper's method over other alternatives.
  • Incremental novelty. DCT action chunks, per-coefficient standardization (EqualSNR), and error-based gating (Gated Memory Policy) are each standard modules.

[1] Xue, Han, et al. "Reactive diffusion policy: Slow-fast visual-tactile policy learning for contact-rich manipulation." arXiv preprint arXiv:2503.02881 (2025).

[2] Hou, Yifan, et al. "Adaptive compliance policy: Learning approximate compliance for diffusion guided control." 2025 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2025.

Questions for authors
  • Routing controls. What do random, fixed-probability, force-threshold, and contact-triggered routing give when using the same coarse and fine experts? Section 4.1.1 describes these comparisons, but no results appear.
  • Can you provide a visualization of the gate value during a few robot rollouts? I would expect higher values during contact and lower values in free space.
  • Gate inputs. Does adding vision, proprioception, or task phase as inputs to gate training helps?
  • What are the inference latency and memory cost of keeping two experts plus the gate?
Reviewer_rFZ9OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SPECTRA uses force history to choose between a coarse policy without force input and a fine policy with force input. The coarse policy predicts low-frequency DCT coefficients, while the fine policy predicts the full action spectrum. After both policies are trained and frozen, a gate learns to select between them using their action-prediction errors on held-out demonstrations. The method uses a pretrained visual encoder and task-trained diffusion policies. Experiments cover six simulation tasks and five real-robot tasks.

Critical strengths
  1. The paper studies when force feedback and finer motion detail are useful. Its experiments with different force inputs and action bandwidths show that the benefit of force feedback depends on the action representation.
  2. The method is evaluated in simulation and on a real robot. It improves average simulation success over always using the fine policy and improves performance on wet-wipe extraction and whiteboard erasing over the reported real-robot baselines.
Critical weaknesses
  1. The paper reports results for fixed policies, but does not report the random, force-threshold, and contact-triggered switching comparisons described in Section 4.1.1. The advantage of the learned gate over these simple switching rules therefore remains unclear.
  2. The external baseline comparison is insufficient. Most comparisons use internal DP/FM variants with different action representations and force inputs. Comparisons with representative force-feedback policies such as RDP and ForceVLA are needed to evaluate overall performance. A comparison with an adaptive method such as FoAR or FA-RDP [1] would help evaluate the selection mechanism. FoAR adjusts force-feature fusion based on predicted contact, while FA-RDP adapts inference frequency and sampling steps. Please analyze how these designs relate to spectral policy selection.

[1] Lifeng Zhuo et al. FA-RDP: A Frequency-Adaptive Reactive Diffusion Policy for Contact-Rich Manipulation. arXiv:2607.28596, 2026. https://arxiv.org/abs/2607.28596

Questions for authors
  1. Please report comparisons with random, force-threshold, and contact-triggered selection using the same frozen coarse and fine policies.
  2. Please compare with representative force-feedback policies such as RDP and ForceVLA, and at least one adaptive method such as FoAR or FA-RDP [1].

[1] Lifeng Zhuo et al. FA-RDP: A Frequency-Adaptive Reactive Diffusion Policy for Contact-Rich Manipulation. arXiv:2607.28596, 2026. https://arxiv.org/abs/2607.28596

3 comments
AuthorsOpenReview ↗

Response to Reviewer rFZ9 (1/2)

We thank the reviewer for recognizing the practical motivation, the interaction between force feedback and action bandwidth, and the improvements demonstrated in simulation and on the real robot. We provide additional routing experiments and clarify the relationship to the suggested methods below.

1. Additional routing-gate comparisons

Table 1 already compares fixed force conditioning, generic force gating, and SPECTRA. With the DP backbone and DCT representation, always using the force-conditioned fine policy achieves 52.9% mean success, while generic force gating achieves 50.0%, compared with 65.8% for SPECTRA across six tasks.

We additionally evaluate random routing and gripper-triggered routing between the coarse and fine policies. The latter uses changes in gripper open/closed state as a contact-related heuristic. We also evaluate bandwidth-only gating between two force-blind action policies to further distinguish the contributions.

TaskRandom routingGripper-triggered routingBandwidth-only gatingSPECTRA
PickCube40.067.562.577.5
PlugCharger15.025.027.532.5
StackCube42.577.575.082.5
Coffee67.560.060.082.5
Threading57.552.577.580.0

Task success rates (%).

Across these five tasks, SPECTRA achieves 71.0% mean success, compared with 44.5% for random routing and 56.5% for gripper-triggered routing. SPECTRA outperforms both alternatives on every task, with improvements of 15–40 percentage points over random routing and 5–27.5 points over gripper-triggered routing.

These results support learning when each expert is useful: introducing two policies and a switching mechanism alone does not recover SPECTRA’s performance. The bandwidth-only comparison further shows that removing force conditioning from both action generators reduces mean success from 71.0% to 60.5%, supporting the combined design.

We will incorporate these results and make the evaluated routing rules explicit, including identifying gripper-triggered selection as an interaction heuristic rather than direct contact detection.

AuthorsOpenReview ↗

Response to Reviewer rFZ9 (2/2)

2. Positioning relative to force-feedback policies and the rationale for controlled comparisons

We appreciate the suggested comparisons and would like to clarify SPECTRA’s specific scientific question: how does the action representation determine when additional sensory information is useful for manipulation?

SPECTRA approaches multimodal policy learning through the structure of the predicted actions. Low-frequency support constrains the policy to smooth motion, while full-spectrum support permits finer corrections. The learned gate jointly selects the sensing inputs and the action subspace used for the next prediction.

The theoretical analyses support this formulation. Appendix C connects sufficiently rapid torque variation to corrections outside the coarse policy’s spectral support under a local compensation model. Appendix D establishes how coefficient-wise scaling balances marginal spectral signal-to-noise ratios and changes the weighting of reconstruction errors. These analyses motivate coupling force conditioning with action representation.

The suggested methods investigate complementary mechanisms:

  • RDP combines slow latent action generation with a fast tactile-conditioned decoder for feedback during chunk execution. SPECTRA selects between independent, complete action generators with different spectral support.

  • ForceVLA performs force-aware feature fusion within a π0-based architecture using pretrained vision-language representations. SPECTRA routes between action policies using supervision from their relative prediction errors.

  • FoAR uses predicted future contact to regulate force-feature fusion and reactive action adjustment. Its native architecture also uses RGB-D-derived point clouds. SPECTRA selects an expert according to relative prediction utility, which is distinct from predicting contact.

  • FA-RDP adapts inference frequency and denoising computation according to estimated multimodality. SPECTRA’s frequency variable is the DCT index within an action chunk. Selecting full-spectrum prediction changes permitted motion detail without itself increasing the feedback rate.

These differences matter when interpreting comparisons. Retaining ForceVLA’s native architecture changes pretrained representations and the generative backbone. Retaining FoAR’s design changes the observation interface and adds reactive action correction. Preserving RDP’s or FA-RDP’s native execution introduces feedback and inference schedules that are central to their methods; constraining those schedules to a common chunk-execution interface would modify their intended mechanisms.

Such comparisons are informative for overall system benchmarking. However, an off-the-shelf ranking would jointly reflect differences in sensing, pretraining, architecture, and execution, and would not isolate the contribution of spectral action structure to multimodal conditioning.

Our experiments therefore prioritize controlled comparisons that directly test this contribution. Table 1 crosses DP/FM backbones, action/DCT representations, and no-force/always-on/gated conditioning under matched demonstration and evaluation settings. The generic-gating baseline already contains two policies, providing a control for introducing another expert and a router.

The additional bandwidth-only ablation then tests whether selecting coarse versus full-spectrum prediction is sufficient when both action generators are force-blind. Its 60.5% mean success, compared with 71.0% for SPECTRA on the same five tasks, complements the modality-only comparison and Figure 3’s force–bandwidth factorial experiment. Together, these results support the contribution at the action-representation level: the usefulness of force conditioning depends on the available action bandwidth, and learning to select their combination improves manipulation performance.

We will sharpen this positioning and clarify SPECTRA’s relationship to feature-fusion, reactive-control, and adaptive-inference methods.

Reviewer_rFZ9OpenReview ↗

Concerns Remain

Thank you for the additional experiments. I would still like to see a simple force-threshold baseline that selects the force-conditioned expert when the measured force exceeds a threshold and the force-blind expert otherwise. I would also like to see comparisons with existing methods such as RDP, ForceVLA and FoAR. I am keeping my current rating and would reconsider it in light of those results.

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

This paper studies force-aware policy learning for contact-rich manipulation. The proposed method, SPECTRA, represents action chunks using the Discrete Cosine Transform DCT and separately trains two diffusion policies: a force-blind coarse policy that predicts only the lower frequencies, and a force-conditioned fine policy that predicts the full action spectrum. A lightweight neural network gate takes the recent force history and selects which policy generates the next action chunk. The method is validated across several simulation and real-robot tasks, showing promising performance in simulation.

Critical strengths
  1. The paper investigates a practical question: force feedback is not always helpful throughout a task, and its benefit depends on whether the policy can express rapid corrections. An interesting finding is that adding force feedback yields no gain when actions are restricted to low frequencies, but improves success rates across multiple tasks when the full action spectrum is available.
  2. The proposed methodology is simple yet effective: training two separate policies (with and without force feedback) at different spectral bandwidths and using a learned gating network to route between them.
Critical weaknesses
  1. The ablation study on the routing gate is missing. Section 4.1.1 explicitly states that learned routing is compared against fixed, random, force-threshold, and contact-triggered selection, but these results do not appear in any table or figure.
  2. In SPECTRA, the gate simultaneously switches between two factors: force vs. no-force conditioning, and coarse policy with low-frequency vs. fine full-spectrum action generation. Cross-evaluations are needed—such as gating between coarse and fine policies under the same observation space (i.e. without force)—to isolate whether the improvement comes from force gating or bandwidth gating. In addition, the choice of the coarse spectral cutoff k requires an ablation study to justify how sensitive performance is to this hyperparameter.
  3. In Table 1 'PegInsertionSide', FM in DCT space without force and with always-on force both outperform SPECTRA and gated FM. While the proposed method does not need to outperform every baseline on every task, this case is worth analyzing: why do both no-force and always-on force outperform gated force on a contact-intensive insertion task?
  4. In Table 4, The real-robot improvements are less significant than those in simulation. It would be helpful to analyze whether real-world force sensor noise or other physical factors limit the effectiveness of the gate and fine policy.
Questions for authors
  1. The ablation study on the routing gate is missing. Where are the results for the fixed, random, force-threshold, and contact-triggered routing baselines mentioned in Section 4.1.1?
  2. To isolate the effect of force vs. action bandwidth, what happens under bandwidth-only gating, i.e. routing between low-frequency and full-spectrum policies when both have no force input? Additionally, how does performance vary with the coarse frequency cutoff k?
  3. In Table 1'`PegInsertionSide', why do both no-force and always-on force outperform the gated policies
  4. Which force/torque sensor is used on the real robot experiments? How does real-world force sensor noise affect gate routing and policy execution?
3 comments
AuthorsOpenReview ↗

Response to Reviewer ZgLP (1/3)

We thank the reviewer for recognizing the practical motivation and effectiveness of our approach. We provide additional ablations below that examine routing quality and separate force conditioning from bandwidth selection.

1. Routing-gate ablations

Table 1 already compares fixed force conditioning, generic force gating, and SPECTRA. With the same DP backbone and DCT representation, no-force prediction, always-on force, and generic force gating achieve mean success rates of 50.0%, 52.9%, and 50.0%, respectively, versus 65.8% for SPECTRA across six tasks. SPECTRA outperforms each of these configurations on every task.

We additionally evaluate random routing and gripper-triggered routing, which select between the coarse and fine policies. The gripper-triggered baseline uses changes in gripper open/closed state as a contact-related routing heuristic. The following table reports these results alongside the bandwidth-only ablation discussed in our next response.

TaskRandom routingGripper-triggered routingBandwidth-only gatingSPECTRA
PickCube40.067.562.577.5
PlugCharger15.025.027.532.5
StackCube42.577.575.082.5
Coffee67.560.060.082.5
Threading57.552.577.580.0

Task success rates (%).

Across these five tasks, SPECTRA achieves 71.0% mean success, compared with 44.5% for random routing and 56.5% for gripper-triggered routing. Its advantage holds on every task. These results support learning which expert to select from interaction history, beyond randomly switching experts or using gripper events as a proxy for when fine control is beneficial. We will incorporate these comparisons and make the evaluated routing rules explicit in the revised experimental description.

AuthorsOpenReview ↗

Response to Reviewer ZgLP (2/3)

2. Separating force conditioning from action bandwidth

The comparisons isolate three distinct choices:

  • Modality-only gating—Table 1(c): select between force-blind and force-conditioned policies, both with full spectral support.
  • Bandwidth-only gating—new ablation: select between low-frequency and full-spectrum policies, both with force-blind action generation.
  • SPECTRA: select between a force-blind low-frequency policy and a force-conditioned full-spectrum policy.

The new bandwidth-only ablation directly tests whether switching between smooth and expressive action generators can explain SPECTRA’s gains without force conditioning in either action policy. As reported in our preceding response, SPECTRA outperforms this variant on all five tasks, improving mean success from 60.5% to 71.0%. The largest improvement is on Coffee, from 60.0% to 82.5%.

These results complement Figure 3, where force conditioning benefits full-spectrum prediction more than low-frequency prediction on four of five tasks. Together, the individual-policy and routing comparisons support the proposed coupling: both the available action bandwidth and the sensory information used to predict those actions matter.

3. Understanding PegInsertionSide

The PegInsertionSide results distinguish the effects of the backbone and the routing configuration. For DCT-space FM, no-force, always-on-force, and generic-gating policies achieve 45.0%, 42.5%, and 37.5%, respectively. For DCT-space DP, the corresponding results are 25.0%, 27.5%, and 27.5%, while SPECTRA reaches 40.0%.

Thus, within the diffusion backbone used by SPECTRA, joint selection improves success by 12.5 percentage points over both always-on force and generic force gating. The strongest absolute result on this task comes from the force-blind FM backbone.

Contact intensity alone does not determine which learned expert is preferable: on this task, adding force improves the DP configuration but reduces the FM configuration. Moreover, the gate’s supervision measures relative action-reconstruction error, while insertion success depends on satisfying precise alignment constraints. A router can therefore make imperfect selections when reconstruction accuracy and closed-loop success are weakly aligned. This provides a plausible explanation for the FM-gating behavior and motivates evaluating routing within a matched backbone and expert configuration.

AuthorsOpenReview ↗

Response to Reviewer ZgLP (3/3)

4. Real-robot sensing and performance

Our OpenArm experiments use the PaXini PX-6AX-GEN3 multidimensional tactile sensor, together with fixed and wrist-mounted cameras. The sensor supports three-axis resultant force and torque output. The manufacturer specifies a minimum recognition force of 0.01 N and repeatability below 0.5% of full scale. We will include the sensor identity and specifications in the implementation details.

Measurement noise can affect both expert selection and force-conditioned action prediction. SPECTRA supplies the gate with an eight-step force history, giving it temporal interaction context rather than basing selection on a single instantaneous force value. Its supervision targets relative expert prediction quality, allowing selection to depend on the usefulness of the resulting action predictions.

The complete real-robot comparison in Table 4 shows that SPECTRA achieves or matches the strongest evaluated baseline on four of five tasks, including force-aware and gated baselines. Wet-wipe extraction improves from 80% to 100%, whiteboard text removal improves from 86.5% to 89.5%, and SPECTRA matches the best performance on Coca-Cola placement and block stacking. These results demonstrate that the approach’s benefits extend to physical execution under real sensing and contact conditions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.