CADeT: Causal-Aware Deformation Transmission for Indirect Robotic Manipulation of Soft Tissue

arXiv:2609.38483 · cs.RO · Submitted 2026-09-29 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: I'm Rosa, and with me are Dev and Taro, guest researcher.

Dev: Today's paper: "CADeT: Causal-Aware Deformation Transmission for Indirect Robotic Manipulation of Soft Tissue".

Rosa: Indirect manipulation of deep-seated deformable anatomy inaccessible to robots is challenging because intervening tissues spatially filter deformation transmission, leading to observational ambiguity between modes.

Dev: First, who's behind it and why it matters.

Title and authors: Rosa: Now we're looking at the title and authors of CADeT, and it points directly to the core challenge they are addressing: indirect manipulation of soft tissue using causal-aware deformation transmission. The authors are Junlei Hu, Dominic Jones Member, IEEE, Pietro Valdastri Fellow, IEEE.

Dev: The title really captures the essence—it’s about getting physical movement through tissue that isn't directly touching the target organ because those tissues spatially filter the deformation transmission pathway.

Taro: I see how this builds on work like Bridge-WA or OGPO, but CADeT focuses specifically on modeling the complex physical mechanism of how that transmission changes based on interaction states, rather than just focusing on trajectory planning.

Rosa: Exactly, and their approach uses an SCM combined with active sensing to infer a latent transmission mode and estimate a state-dependent adhesion Jacobian online during normal manipulation.

Dev: That means they aren't just guessing the tissue behavior; they are actively learning the physics of that interaction on the fly by using control actions to update their belief about the mode.

Taro: It’s a sophisticated way to move away from static models when dealing with heterogeneous tissues, which is where most traditional robots struggle.

Rosa: And they validate this whole system through simulation and experiments on both phantom and ex vivo porcine tissues using the da Vinci research kit or dVRK setup.

Dev: I'm curious if this works outside of that lab setting; Rosa, how long do you think this kind of adaptive control loop could maintain stability when deployed in a more open surgical environment?

Taro: That’s the million-dollar question for autonomy; if it can handle the uncertainty in a known setup, its ability to manage unpredictable real-world scenarios is what we really need to test.

The paper's summary: Rosa: To summarize CADeT, the authors propose a framework that builds an SCM to represent how the latent transmission mode modulates the proxy-to-target deformation pathway, which then informs mode-conditioned observation models for a Bayesian update to maintain this belief.

Dev: So, they use that SCM to have a structural basis for understanding how different physical modes affect what the robot sees and feels during manipulation.

Taro: They are essentially creating an internal representation of the tissue's physical behavior—whether it's decoupled or blocked—which is not immediately obvious from just looking at the tissue deformation.

Rosa: When ambiguity persists between modes, such as Decoupled and Blocked, the framework selects an additional probing action specifically to improve that mode distinguishability before proceeding with control.

Dev: That’s a clever way to resolve observational ambiguity by treating information gain as a priority alongside trajectory tracking when the system is uncertain.

Taro: I think this active sensing component is what makes it useful in situations where passive observation fails, which is exactly what you see when things get messy during surgery.

Rosa: And for the actual control part, they integrate this mode belief and a Gaussian process estimate of the adhesion Jacobian within a model predictive control framework to regulate target deformation through the proxy organ.

Dev: So, if they believe it's in coupled mode, they use that learned Jacobian to accurately predict how much force will transmit, which is essential for precise control in that MPC setting.

Taro: That coupling of inference and control planning seems like a very thorough way to handle the physical reality of soft tissue manipulation.

The paper's improvements: Rosa: The paper highlights two main areas for improvement, first by showing that they can distinguish between interaction-state changes like sticking or sliding just from force or motion measurements, which is a nice feature.

Dev: But the real focus seems to be on how they handle the ambiguity where passive observations fail to uniquely indicate the underlying physical interaction mode, quantified using Shannon entropy H(τt Ht).

Taro: That uncertainty metric, where H(τt Ht) > zero point two five nat in their experiments, gives us a concrete way to measure when the system is genuinely confused between modes.

Rosa: To resolve that confusion, they show how different probing actions yield different expected responses under the Decoupled versus Blocked modes, for example, Y(zero) t (δu) ≈ µn (δu) zero versus Y(two) t (δu) ≈ λbµn (δu) zero.

Dev: That comparison of expected responses is what allows them to select the probing action that gives the most information gain, which is a very specific and actionable way to improve model confidence.

Taro: This moves beyond just picking any random action; it’s an intelligent selection process designed specifically to resolve the physical ambiguity in real time.

Rosa: They also showed how they can incorporate this mode belief and the learned Jacobian into a belief-aware model predictive controller for indirect target-shape control, which is how they translate their inference into actual physical movement.

Dev: That integration means the system isn't just guessing based on a single observation; it's planning its next steps based on what it *thinks* the physics are doing right now.

Conclusion: Rosa: To wrap up CADeT, they’ve shown how you can use an SCM and active sensing to infer the transmission mode while using a belief-aware MPC to perform indirect shape control, which is a significant step for manipulating deep anatomy.

Dev: The main implication is that this framework allows for high-precision indirect shape control even when tissue interaction physics are highly uncertain, provided you have the right active sensing loop running.

Taro: I think the real impact here is demonstrating how to build systems that can maintain stable manipulation in environments where the physical coupling state isn't perfectly known beforehand.

Rosa: Indeed, and they validate this on both phantom and ex vivo porcine tissues using the da Vinci research kit, showing transmission-mode identification works under different interaction conditions.

Dev: If this can be made robust enough for longer periods in a clinical setting, it means we could see much more reliable robotic assistance for delicate procedures without needing perfect pre-operative knowledge of tissue structure.

Taro: I'm just thinking that the ability to probe and adapt based on real-time ambiguity resolution is what really matters when dealing with the unpredictable nature of biological systems.

Rosa: So, CADeT gives us a powerful tool for moving beyond rigid models in soft tissue manipulation, and it’s definitely something worth watching as we look at next steps for this research.

Junlei Hu, Dominic Jones Member, IEEE, Pietro Valdastri Fellow, IEEE

University of Leeds

cs.RO

Submitted: 2026-09-29

Updated: 2026-09-29

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 77/100

The gist: Indirect manipulation of deep-seated deformable anatomy inaccessible to robots is challenging because intervening tissues spatially filter deformation transmission, leading to observational ambiguity

Key concepts

Structural Causal Models (SCM): An SCM is a mathematical structure used to represent the causal relationships between different variables. In this work, it models how the latent transmission mode influences how deformation travels from a robot's proxy organ to the actual target tissue, providing the necessary structure for making predictions.
Transmission Modes: These are distinct ways soft tissue deformation is transmitted between organs. The framework identifies three main modes: Decoupled (no mechanical link), Coupled (mechanically linked), and Blocked (external constraint restricts transmission). Identifying the correct mode is crucial for accurate control.
Mode Belief: This represents the robot's current statistical understanding of which transmission mode is active. The system maintains this belief through Bayesian updates based on tissue responses to control actions. If the ambiguity between modes remains high, the system triggers probing actions to resolve it.
Adhesion Jacobian: This is a state-dependent mathematical function that describes how much the target organ deforms in response to changes in its internal state, specifically when organs are coupled. Estimating this Jacobian online allows the robot to calculate necessary control inputs for indirect shape control.

Terminology

Adhesion Jacobian — 15× in this paper · explained in 1 episode
Accuracy — 13× in this paper · explained in 1 episode
Ablation — 11× in this paper · explained in 1 episode
Information Gain — 10× in this paper · explained in 2 episodes
Robustness — 4× in this paper · explained in 5 episodes
Causal Inference — 3× in this paper · explained in 4 episodes
International Conference on Robotics and Automation — 3× in this paper
Reinforcement Learning — 3× in this paper · explained in 4 episodes
Structural Causal Model (SCM) — 3× in this paper · explained in 4 episodes
Active Perception — 2× in this paper · explained in 2 episodes
Calibration — 2× in this paper · explained in 7 episodes
Core — 2× in this paper · explained in 2 episodes
Deep Reinforcement Learning — 2× in this paper · explained in 2 episodes
Gaussian Process Regression — 2× in this paper
Model Predictive Control (MPC) — 2× in this paper · explained in 5 episodes
Transfer Learning — 2× in this paper · explained in 10 episodes
Action Space — 1× in this paper · explained in 1 episode
Benchmark — 1× in this paper · explained in 2 episodes
Causal Reasoning — 1× in this paper · explained in 2 episodes
Consistency — 1× in this paper · explained in 3 episodes
Physical Constraints — 1× in this paper · explained in 7 episodes
Regularization — 1× in this paper · explained in 3 episodes
Segmentation Masks — 1× in this paper · explained in 2 episodes
Structural Causal Models — 1× in this paper

Summary

Indirect manipulation of deep-seated deformable anatomy inaccessible to robots is challenging because intervening tissues spatially filter deformation transmission, leading to observational ambiguity between modes. This work proposes CADeT, a causal-aware framework that integrates Structural Causal Models (SCM) with active sensing to infer a latent transmission mode and estimate a state-dependent adhesion Jacobian online for indirect target-shape control.

How it works

An SCM is constructed to represent how the latent transmission mode modulates the proxy-to-target deformation pathway, providing the structural basis for mode-conditioned observation models used in a Bayesian update to maintain the mode belief.
The framework continuously updates this mode belief using tissue responses induced by normal control actions. When ambiguity persists between modes like Decoupled and Blocked, an additional probing action is selected to improve mode distinguishability.
A deformation transmission control (DTC) strategy uses the inferred mode belief and learned coupling model to regulate target deformation through the proxy organ, formulated within a model predictive control (MPC) framework.

Key Components of the Framework

(Note: The paper enumerates several key components, which are summarized below.)

An SCM-based transmission-mode inference framework that uses active sensing with additional probing actions to infer the latent transmission mode online.
A DTC scheme for indirect shape control that integrates the mode belief and a Gaussian process (GP) estimate of the adhesion Jacobian within a model predictive control framework.
Experimental validation on the da Vinci research kit (dVRK) using both phantom and ex vivo tissues to demonstrate transmission-mode identification and indirect shape control under different interaction conditions.

Modeling Transmission Dynamics

The system is modeled by a hybrid transmission function, where the target dynamics are formulated as:

(Note: The paper details the three modes.)

In the Decoupled Mode (τt = 0), adhesion between the proxy and target is insufficient or broken, resulting in T (• τt = 0) ≈ 0.
In the Coupled Mode (τt = 1), organs are mechanically linked, and transmission is modeled as T (• τt = 1) ≈ Jadh(gt; θ) · (gt+1 − gt), where Jadh is a state-dependent adhesion Jacobian.
In the Blocked Mode (τt = 2), an external constraint locally restricts deformation transmission along the current manipulation direction, resulting in T (• τt = 2) ≈ 0, despite continued proxy deformation.

Handling Observational Ambiguity

The paper addresses the ambiguity where passive observation fails to distinguish between modes, such as when the observed tissue deformation may no longer uniquely indicate the underlying physical interaction.

This ambiguity is quantified using Shannon entropy H(τt Ht), with uncertainty indicated by "H(τt Ht) > Hth," where Hth = 0.25 nat in experiments.
To resolve this, the robot applies an intervention action do(δu), and the expected response under different modes is compared: Y (0) t (δu) ≈ µn (δu) 0 for Decoupled Mode, versus "Y (2) t (δu) ≈ λbµn (δu) 0, 0 < λb < 1" for Blocked Mode.

Active Sensing and Control Objectives

The control objective is to compute a finite-horizon control sequence that drives the target toward a desired configuration pd under the current conditions, formulated as:

(Note: The paper details the MPC formulation.)

The cumulative cost is defined as J(ut:t+T −1) = T X−1 k=0 pˆt+k+1 − pd 2 Q + ut+k 2 R.
The optimal control sequence is found by minimizing the expected cost under the current mode belief: u∗ t:t+T −1 = arg min ut:t+T−1 Eτ∼bt [J(ut:t+T −1)].

Probing Action Selection

To maximize information gain about the latent transmission mode, the robot selects a probing action δu that maximizes mutual information I(τt; Yt Ht, do(δu)).

(Note: The paper details the optimization strategy.)

The expected information gain is approximated by maximizing the entropy of the marginal predictive distribution H(Yt δu).

Improvements for AI systems

Based on the provided research paper, here are specific improvements that can be made to existing AI systems, particularly in robotic manipulation and control:

) Improved System Capability 1: Causal-Aware Deformable Object Manipulation (CADeT) for RAMIS

The core improvement is moving from purely data-driven or model-free methods to a framework that explicitly reasons about the physical mechanism of interaction.

This improved system can perform indirect manipulation of deep-seated, deformable anatomy (like organs) by acting as a sophisticated surgeon's assistant. Specifically, it can:

Distinguish between two critical failure modes—the target being mechanically decoupled from the proxy (Decoupled Mode) versus the target being physically constrained by external anatomy (Blocked Mode)—even when both result in nearly identical visual observations (e.g., stationary target motion).
Update its internal knowledge of how the tissue is coupled in real-time based on its own actions and active probing, leading to a robust understanding of the underlying physics rather than relying on potentially mismatched global models.

) Improved System Capability 2: Belief-Aware Model Predictive Control (MPC) for Indirect Shape Control

The system can perform high-precision, goal-oriented control of soft organs by leveraging its inferred physical state. Specifically, it can:

Plan a sequence of proxy motions that explicitly account for the uncertainty in the tissue coupling mode. If it believes the organs are coupled (Coupled Mode), it uses a learned adhesion Jacobian to predict and minimize target tracking error accurately. If uncertainty is high, it plans conservatively or initiates probing actions to gain more certainty before committing to a move.
Execute shape control even when the target is partially occluded, by dynamically adjusting its control authority based on the inferred mode belief (i.e., reducing predicted transmission if coupling is suspected to be lost).

) Improved System Capability 3: Active Sensing and Information-Guided Probing for Model Refinement

The system can proactively seek out information that it needs to distinguish between ambiguous states. Specifically, it can:

Automatically select the most informative physical action (probing action) to execute when its current belief about the system state is uncertain (i.e., when two modes look the same). This selection is based on maximizing the expected information gain about which mode is active, rather than just picking a random move.
Improve its confidence in its learned physical models (like adhesion Jacobians) by intelligently choosing probing directions that produce the largest possible difference between competing physical scenarios.

) Improved System Capability 4: Robustness to Early Errors via Confidence Gating

The system can maintain high performance and prevent catastrophic failures if it makes a mistake early in the manipulation sequence. Specifically, it can:

Ignore or gate early observations that contradict its current belief (e.g., misclassifying a mode) when updating its physical models (like the GP coupling model). This prevents transient errors from corrupting the long-term, stable understanding of the tissue dynamics.