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

Junlei Hu; Dominic Jones; Pietro Valdastri

cs.ROSep 29, 2026

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

Authors: Junlei Hu, Dominic Jones, Pietro Valdastri

Organizations: STORM Lab, Institute of Autonomous Systems and Sensing (IRASS), School of Electronic and Electrical Engineering, University of Leeds, Leeds, UK

Abstract

Indirect manipulation of deep-seated deformable anatomy inaccessible to the robot is challenging in robot-assisted minimally invasive surgery (RAMIS) because intervening tissues spatially filter deformation transmission. Passive observations can be ambiguous because the Decoupled and Blocked modes may produce similar motion responses. We propose CADeT, a causal-aware deformation transmission framework that integrates structural causal model (SCM) with active sensing to infer a latent transmission mode and estimate a state-dependent adhesion Jacobian online. During normal manipulation, control actions update the mode belief; when ambiguity persists, an additional probing action is selected to improve mode distinguishability. The mode belief and learned Jacobian are incorporated into a belief-aware model predictive controller for indirect target-shape control. Validation in simulation and on the da Vinci research kit (dVRK), using phantom and ex vivo porcine tissues, shows higher mode-identification accuracy and faster shape-error convergence than the evaluated model-free and model-based baselines. These results show that active sensing improves mode identification and indirect deformation control under the evaluated conditions.

Figures & tables

Fig. 1: Conceptual illustration of indirect deformation transmission. High-frequency robotic inputs at the proxy surface (stomach) are attenuated by viscoelastic strain-energy dissipation, producing a low-frequency, globalized deformation whose transmission to the target organ (pancreas) depends on the latent transmission mode.

Fig. 2: Kinematic configuration for indirect manipulation in RAMIS. The robot actuates the proxy tissue while the target state is observed visually, with an external constraint representing the Blocked Mode.

Fig. 3: structural causal model of indirect deformation transmission. The control action $\mathbf{u}_{t}$ drives proxy deformation $\Delta\mathbf{g}_{t}$ , while the latent transmission mode $\tau_{t}$ gates the proxy-to-target transmission pathway.

Fig. 4: Conceptual illustration of Passive Observation and Active Sensing.

Panel of Figure 5 — Fig. 5: Representative simulation results for the three transmission modes: (a) Decoupled, (b) Coupled, and (c) Blocked. The image sequences show stomach-proxy manipulation and the pancreatic-target response. The stacked area plots show the mode belief $b_{t}(\tau)$ , while the red curve shows the mean target-shape error. Target deformation and error reduction are evident in the Coupled Mode, whereas target motion remains limited in the Decoupled and Blocked modes.

Fig. 6: Confusion matrices comparing mode-identification accuracy between the no-additional-probing and the full CADeT framework across the Decoupled, Coupled, and Blocked modes.

Panel of Figure 7 — Fig. 7: Evolution of the mode-belief entropy for the no-probing variant and the full CADeT framework across the three transmission modes. Lower entropy indicates higher confidence in the inferred mode.

Method	Mode ID. Accuracy (%)	Blocked Recall (%)	Belief Conv. Time (s)	Probe Count
Random probing	$94.0$	$90.0$	$14.9\pm 1.5$	$15.5\pm 4.1$
CADeT	$\mathbf{99.3}$	$\mathbf{98.0}$	$\boldsymbol{13.2\pm 0.4}$	$\boldsymbol{6.7\pm 1.1}$

TABLE I: Ablation of the probing policy.

Fig. 8: Representative mode-belief trajectories on the 2-simplex under random probing in the Coupled Mode. The vertices denote certainty in the Decoupled ( $\tau=0$ ), Coupled ( $\tau=1$ ), and Blocked ( $\tau=2$ ) modes, and the colour gradient indicates temporal progression from the initial uniform belief (green circle) to the final belief (red square). (a) and (b) show rapid convergence, whereas (c) shows a slower trajectory with temporary ambiguity.

$\boldsymbol{\lambda}_{\mathrm{b}}$	Mode Accuracy (%)	Blocked Recall (%)	B $\rightarrow$ D Rate (%)	Belief Conv. Time (s)
$0.2$	$98.0$	$94.0$	$4.0$	$15.4\pm 0.5$
$0.4$	$\boldsymbol{99.3}$	$\boldsymbol{98.0}$	$\boldsymbol{0.0}$	$\boldsymbol{13.2\pm 0.4}$
$0.6$	$98.7$	$96.0$	$4.0$	$13.7\pm 0.3$
$0.8$	$95.3$	$86.0$	$14.0$	$17.9\pm 2.1$
$0.95$	$91.3$	$82.0$	$18.0$	$20.2\pm 2.4$

TABLE II: Sensitivity of mode inference to the blocked-response attenuation factor

\lambda_{\mathrm{b}}

Panel of Figure 11 — Fig. 9: Shape-control robustness to early mode misclassification under different GP update strategies.

Variant	Mode Inference	GP Coupling	Belief Weighting	Success Rate	Task Time (s)	Average DTR
No mode conditioning	Yes	Yes	None	$84.0\%$	$37.7\pm 4.2$	$7.54\%$
Hard-mode MPC	Yes	Yes	Hard	$90.0\%$	$45.1\pm 5.1$	$8.67\%$
Jacobian coupling	Yes	No	Soft	$86.0\%$	$38.6\pm 3.7$	$7.82\%$
Full CADeT	Yes	Yes	Soft	$\mathbf{98.0\%}$	$\mathbf{30.1\pm 2.7}$	$\mathbf{11.26\%}$

TABLE III: Component ablation of mode conditioning, belief weighting, and GP-based coupling estimation.

Panel of Figure 13 — Fig. 10: Deformation transmission ratio achieved by CADeT and the baseline controllers. A higher ratio indicates more effective transmission of proxy deformation to the target.

Method	Success Rate	Conv. Speed	Average DTR
Jacobian-based [ 45 ]	$42.0\%$	$1.0\times$	$4.25\%$
GP-WRM [ 4 ]	$68.0\%$	$1.17\times$	$5.27\%$
Online GPR [ 18 ]	$75.0\%$	$1.20\times$	$5.22\%$
FEM-based [ 6 ]	$75.0\%$	$0.61\times$	$3.92\%$
CADeT (Ours)	$\mathbf{98.0\%}$	$\mathbf{2.2\times}$	$\mathbf{11.26\%}$

TABLE IV: Comparison with existing manipulation methods under the tested coupling conditions.

Fig. 11: Experimental setup for the RAMIS validation on the dVRK platform, detailing the phantom and ex vivo environments.

Fig. 12: SAM2–DINOv3 perception pipeline for marker-free tissue segmentation, feature tracking, and shape-state estimation.

Panel of Figure 17 — Fig. 13: The left columns show time-series snapshots of the surgical scene, with the stomach proxy represented by the red boundary and red feature points, the pancreatic target by the green boundary and green feature points, and yellow points denoting the desired target configuration. The right columns show the mode belief $b_{t}(\tau)$ as a stacked area chart together with the mean target-shape error $e_{t}$ shown by the red solid line.

Panel of Figure 18 — Fig. 14: The left columns show time-series snapshots of the surgical scene, with the colon proxy represented by the red boundary and red feature points, the mesentery target by the green boundary and green feature points, and yellow points denoting the desired target configuration. The right columns show the mode belief $b_{t}(\tau)$ as a stacked area chart together with the mean target-shape error $e_{t}$ shown by the red solid line.

Case	Tissue	Test Condition	Mode Identified	Initial Error (mm)	Final Error (mm)
Pancreas (a)	Ex vivo	Decoupled	$\checkmark$	$39.8$	$36.6$
Pancreas (b)	Ex vivo	Blocked	$\checkmark$	$57.6$	$2.8$
Pancreas (c)	Phantom	Coupled	$\checkmark$	$45.3$	$2.3$
Pancreas (d)	Phantom	Decoupled	$\checkmark$	$52.1$	$53.5$
Mesentery (a)	Ex vivo	Coupled	$\checkmark$	$36.8$	$2.7$
Mesentery (b)	Ex vivo	Decoupled	$\checkmark$	$43.8$	$43.8$

TABLE V: Quantitative results of the physical RAMIS experiments.

Explore similar work

Deformable Object Manipulation under Partial Observability via Real-Time Full-Shape Estimation
Sep 9, 2026Kosar Behnia, Ville Kyrki, Gokhan AlcanDeformable Linear Objects Statistical Shape Model
Faculty of Engineering and Natural Sciences, Tampere Univ., Finland. · Electrical Engineering and Automation Department, Aalto Univ., Finland.
Shaft-integrated Force Sensing with Transformer-based Dynamics Compensation for Telesurgery
May 29, 2026Shuyuan Yang, Grant Boone, Timo Markert +4Surgical Robots Ground Reaction Forces
Department of Electrical, Computer, and Systems Engineering, Case Western Reserve University, Cleveland, OH 44016, USA · Department of Mechanical and Aerospace Engineering, Case Western Reserve University, Cleveland, OH 44106, USA · Resense GmbH, Germany +3
Learning Robust Dexterous In-Hand Manipulation from Joint Sensors with Proprioceptive Transformer
May 20, 2026Senlan Yao, Chenyu Yang, Jaehoon Kim +2Proprioception
Soft Robotics Laboratory, ETH Zürich, Switzerland

Sep 9, 2026, cs.ROJ/K move · Enter open · S save

Deformable Object Manipulation under Partial Observability via Real-Time Full-Shape Estimation

Kosar Behnia, Ville Kyrki, Gokhan Alcan

Faculty of Engineering and Natural Sciences, Tampere Univ., Finland. · Electrical Engineering and Automation Department, Aalto Univ., Finland.

Deformable Linear Objects Statistical Shape Model Partial Observability Underactuated Dynamics Conditional Variational Autoencoder

Open paper AskPDF arXiv

Manipulating deformable objects (DOs) is challenging due to their high-dimensional state space, underactuated dynamics, and partial observability. In this paper, we propose cRVAE, a lightweight conditional recurrent variational autoencoder that estimates the full DO state from only partial corner-node observations during inference. The resulting model is used as the forward model in a receding-horizon optimal control framework for obstacle-aware collaborative DO manipulation. In simulation on rope and fabric, cRVAE estimates the full DO state from the available corner-node measurements alone, matching the accuracy of a parameter-identified XPBD model. At inference it uses no physical parameters as model inputs and performs no online parameter identification. It also runs approximately 350 times faster on the rope and over 1500 times faster on the fabric per forward pass, keeping horizon-based planning within the 100 ms control budget where XPBD exceeds it already at short horizons. Full-shape estimation from corner sensing at in-loop speed is what makes the model deployable on hardware, which we demonstrate on a Unitree Go2 robot.

Abstract

Explore similar work

Deformable Object Manipulation under Partial Observability via Real-Time Full-Shape Estimation

Shaft-integrated Force Sensing with Transformer-based Dynamics Compensation for Telesurgery

Learning Robust Dexterous In-Hand Manipulation from Joint Sensors with Proprioceptive Transformer