CADeT

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

arXiv preprint · September 2026

Junlei Hu, Dominic Jones, Pietro Valdastri

STORM Lab · University of Leeds

Manipulate an accessible tissue to move a target that the robot cannot grasp directly. CADeT uses tissue responses and targeted probing to infer how deformation is transmitted, then adapts indirect shape control to that interaction.

A robot grasps the stomach to transmit deformation to the pancreas; the illustration shows attenuation through the proxy tissue and an external constraint.
Indirect manipulation through a proxy organ: the robot acts on the stomach while tracking the pancreatic target. Figure 1 from the paper.

The same target response can have different causes

A target that barely moves may be disconnected from the manipulated tissue, or attached but blocked by its surroundings. CADeT maintains a belief over three transmission modes:

Mode Interaction Target response
Decoupled Ineffective proxy–target coupling Little transmitted motion
Coupled The proxy transmits deformation The target responds to proxy motion
Blocked The coupled target meets an external constraint Motion is restricted
  1. Observe and infer. Track the proxy and target tissues, updating the transmission-mode belief from their responses.
  2. Probe when ambiguous. Choose an additional action whose predicted responses help distinguish competing modes.
  3. Control through the proxy. Combine the belief with an online estimate of the adhesion Jacobian in model predictive control.

Method details: Sections IV–VI of the paper.

Results at a glance

In the paper's simulation evaluation, full CADeT achieved:

99.3%mode-identification accuracy
98.0%manipulation success
30.1 smean task time (± 2.7 s)

Simulation results: 149/150 trials correctly identified. Manipulation statistics are reported separately in Table III. Source: Section VII, Tables I and III.

Physical experiments used the da Vinci Research Kit with tissue phantoms and ex vivo porcine tissues. In the three nominal reachable Coupled cases, final mean shape errors were 2.1–2.7 mm (Table V).

Scope. Correctly identifying coupling does not make every target shape reachable. The experiments include disconnected, constrained and reachability-limited cases. The method assumes a predefined grasp and does not provide a formal tissue-safety guarantee.

Experiment videos

Each clip contains one complete demonstration with its original plots and playback speed. Use the controls to pause, seek or view fullscreen.

Each excerpt is 25 seconds long. On-screen experimental time can differ from video playback time.

Pancreatic manipulation through the stomach

The stomach is the accessible proxy; the pancreas is the indirectly manipulated target. Paper: Section VIII-B.

Ex vivo · decoupledThe porcine stomach and pancreas are not attached. Probing identifies ineffective transmission.
Ex vivo · constraint engagementAttached tissues encounter a surrounding constraint; the belief responds as the interaction changes.
Phantom · coupledThe attached stomach transmits deformation to the pancreas model.
Phantom · decoupledThe stomach and pancreas models are not attached.

Mesenteric manipulation through the colon

The robot manipulates the colon to control the attached mesentery. Phantom trials use a simplified intestinal model. Paper: Section VIII-C.

Ex vivo · coupledThe robot manipulates the colon to deform the attached porcine mesentery.
Ex vivo · decoupledThe porcine colon and mesentery are not attached.
Phantom · coupledA simplified intestinal phantom demonstrates effective deformation transmission.
Phantom · constraint engagementThe attached target encounters a surrounding constraint during manipulation.

Simulation: three transmission conditions

Controlled attachment and blocking conditions illustrate Decoupled, Coupled and Blocked interactions.

DecoupledNo attachment transmits stomach motion to the pancreatic target.
CoupledThe attached proxy transmits motion to the target.
Constraint engagementThe coupled target interacts with an external constraint.

Citation

@misc{hu2026cadet,
  title = {{CADeT}: Causal-Aware Deformation Transmission for
           Indirect Robotic Manipulation of Soft Tissue},
  author = {Junlei Hu and Dominic Jones and Pietro Valdastri},
  year = {2026},
  eprint = {2609.38483},
  archivePrefix = {arXiv},
  primaryClass = {cs.RO},
  url = {https://arxiv.org/abs/2609.38483}
}