acceptodds
Withdrawn from ICLR 2027

Touch2Predict: Learning Physical Properties from a Single Tactile Interaction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Estimating physical properties of deformable objects is essential for predictive robotic manipulation, but remains difficult in contact-rich interactions where visual deformation is small, localized, and often occluded. Tactile sensing provides direct measurements of mechanical response, yet these observations are highly sensitive to uncertain robot–object contact configuration, coupling pose and material estimation. We introduce Touch2Predict, a structured tactile inverse-modeling framework that separates these two factors according to their physical roles. Contact-establishment cues guide a low-dimensional derivative-free search over contact-relevant pose variables, while the full spatiotemporal pressure sequence drives adjoint-based material refinement through differentiable contact simulation. Alternating the two stages progressively recovers the interaction-specific contact state and object-level physical properties. Across eight synthetic scenes with translational, rotational, and coupled pose uncertainty, \method recovers Young's modulus with an average per-scene median relative error of 6.05%. Real-robot experiments further demonstrate its practical feasibility, with the modulus identified for a test object within 12.4% of an independent reference measurement. Together with qualitative predictions under unseen interactions, these results highlight the potential of tactile inverse physics to recover transferable physical models for understanding and predicting physical interactions.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Tactile Sensing, Material Parameter Identification, Contact Pose Uncertaint, Differentiable Simulation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_Yc7tOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Touch2Predict, a tactile-based inverse physics framework that estimates the physical properties of deformable objects from a single robotic probing interaction. It alternates between contact pose refinement and material parameter optimization using differentiable physics simulation to recover properties such as Young’s modulus and predict unseen interactions.

Critical strengths
  1. Well-motivated decomposition of contact pose and material estimation. The paper addresses the important challenge of pose–physics ambiguity in tactile-based material identification. By separating pose refinement using contact-establishment cues from material optimization using full pressure observations, the proposed framework provides a physically motivated approach to a difficult inverse problem. This structured formulation is the paper's primary methodological contribution and offers a potentially useful direction for contact-rich robotic perception.

  2. Strong simulation results with encouraging real-world validation. The method is evaluated across eight synthetic scenes with different types of pose uncertainty, supported by comparisons against several baselines and ablation studies. The real-robot experiments further demonstrate the feasibility of identifying material properties from a single tactile interaction. These results provide meaningful empirical support for the effectiveness of the proposed optimization framework.

Critical weaknesses
  1. Strong dependence on accurate object geometry and structural assumptions. The proposed method assumes that the object geometry is known and fixed, while only contact pose and material parameters are optimized. However, for deformable objects with complex structures (e.g., hollow or thin-walled objects), the observed deformation and tactile response depend not only on Young's modulus but also on structural properties such as wall thickness and internal geometry. Errors in the reconstructed geometry could therefore be compensated for by incorrect material estimates. The paper does not systematically investigate this geometry-material ambiguity or evaluate robustness to geometric inaccuracies, which raises concerns about the reliability of material identification in real-world scenarios.

  2. Lack of convincing downstream task evaluation. Although the paper motivates material identification as a means to improve predictive robotic manipulation, it does not demonstrate how the recovered physical properties benefit actual manipulation, planning, or control tasks. The generalization experiments are primarily limited to qualitative predictions under unseen interactions, and the real-robot evaluation involves only two objects. Without quantitative evaluations of predictive generalization or downstream task performance, it remains unclear whether the estimated physical models provide meaningful practical advantages beyond fitting the observed tactile responses. This substantially weakens the paper's claims regarding the utility and transferability of the proposed framework.

Questions for authors
  1. How are volumetric FEM meshes constructed for hollow objects such as the Cup, especially when internal geometry is not visible? How sensitive is the estimated Young's modulus to errors in mesh geometry or wall thickness? How does the method distinguish material stiffness from structural stiffness using only a single tactile interaction?

  2. Can the authors provide quantitative prediction errors for the held-out interactions, rather than only qualitative comparisons?

  3. How does the method perform when pose uncertainty extends beyond the optimized low-dimensional pose parameters?

  4. How do excluded simulation failures affect the reported results and overall success rate?

  5. Can the authors involve more downstream tasks to illustrate the importance of the whole estimation framework?

Reviewer_WMKXOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Touch2Predict, a tactile inverse-modeling framework for estimating the Young's modulus of a deformable object from a single probing interaction under uncertain robot-object pose. It alternates between pose refinement using structured contact cues and material estimation using the full tactile pressure response. Across eight synthetic scenes, the method achieves lower median modulus errors than the main baselines. The paper also includes preliminary experiments on two real-world objects.

Critical strengths

The paper recognizes that the coupling between pose and material estimation is an important practical problem.

Critical weaknesses

The experimental setup is my biggest concern. The main results come from simulations using ground-truth meshes and noise-free contact forces. The authors test Touch2Predict in exactly the same simulation environment, which means it effectively has access to a perfect world model. They mention errors in calibration, depth estimation, and registration, but do not consider these errors in their experiments.

The only quantitative real-world result is the 12.4% error for the yellow egg, which does not make sense to me. Touch2Predict should find the pose and Young's modulus that best match the simulated contact forces to the measured ones. However, Figure 5 shows that the simulated forces are consistently lower than the measured forces. This suggests that the system should increase its estimate of Young's modulus, which would likely make the error larger than 12.4%.

Questions for authors
  1. For Banana_r, the pose barely improves, yet the modulus error is only 3.8%. For Torus, the pose is recovered accurately (1.35mm), but the modulus error is 26.9%. This contradicts the proposed ambiguity that an incorrect pose paired with an incorrect modulus can still perfectly fit the observations. How do you explain these results?
  2. Why is the blue ball's modulus not reported?
  3. Minor issue: inconsistent capitalization of Touch2Predict / TOUCH2PREDICT.
Reviewer_JoAgOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Touch2Predict estimates an object's physical properties, like stiffness, directly from a single tactile probing interaction. The core idea is to tackle "pose-physics compensation," a tricky issue where a slightly wrong contact pose can easily fool the system into guessing the wrong material stiffness. To get around this, the method alternates between two steps: first, it uses contact timing and pressure shifts to lock in the exact pose, and then it uses a differentiable physics simulator to dial in the actual material parameters. It works quite well in practice, hitting around a 6% error in simulations and 12% on a real robot, while still predicting how the object will deform in unseen poses. The main catch, though, is that this alternating optimization is too slow for real-time robotic control, and the system really struggles if the object is too soft for the sensors to pick up.

Critical strengths

Rather than relying blindly on end-to-end deep learning, this paper provides a profound analysis of the core ambiguity inherent in tactile inverse physics, specifically, the coupled compensation between pose and material parameters—and proposes a decoupling strategy rooted in physical mechanisms, offering a solid theoretical foundation. The paper employs an ingenious alternating optimization scheme and a well-designed objective function: instead of optimizing pose directly using raw pressure residuals (which can be misleading due to the influence of material parameters), the authors introduce a "Structured Pose Objective" that incorporates contact evidence, temporal information, and the spatial center of pressure. This design significantly enhances the robustness of pose estimation, even when material parameters are inaccurate.

The method is built upon an advanced GPU-accelerated differentiable contact dynamics framework (StiffGIPC/DiffIPC), which enables the computation of adjoint gradients for complex, non-smooth contact dynamics—including collision detection and friction—representing a high technical barrier.

Critical weaknesses

Despite the use of GPU acceleration, the alternating optimization process—consisting of multiple rounds of CMA-ES evaluation followed by L-BFGS optimization via the adjoint method—remains computationally expensive. As shown in Table 11 of the appendix, even with low-dimensional parameter settings, a single complete identification task requires between 20 and 100 minutes of cumulative GPU computation time. Consequently, the method is currently suitable only for offline parameter calibration or slow-speed operations, making it ill-suited for highly dynamic, real-time closed-loop control.

The method relies on the availability of a "visually reconstructed mesh" of the object and a rough initial pose. If the initial mesh contains significant topological errors or geometric deviations, the forward predictions from the differentiable simulator will exhibit systematic bias, potentially causing the entire alternating optimization process to become trapped in a local optimum.

Although the appendix mentions the possibility of simultaneously identifying Young’s modulus and Poisson’s ratio (or yield stress), these parameters may exhibit high collinearity in their influence on tactile pressure—especially when dealing with real-world, noisy tactile signals. As the paper primarily demonstrates multi-parameter identification using synthetic data, further validation is required to assess the method's stability and identifiability in complex, real-world scenarios.

The paper also acknowledges in its conclusion that "real-robot evaluation remains limited," as it was validated using only two simple geometric objects: an egg and a sphere. Furthermore, regarding the experiment with the blue sphere, the authors admit that "discrepancies in force magnitude and spatial distribution remain" when predicting new poses.

Overall, I'm leaning toward rejection, but I don't have a strong stance on this. If the authors can adequately address my concerns, or if other reviewers see strong merits in the work, I'm open to reconsidering my recommendation and supporting acceptance.

Questions for authors

See above

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

Related papers

Discussion (0)

Sign in to comment.