Force-informed policy
Contact force and robot state jointly condition continuous control.
Contact-rich control · Flexible endoscopy
CAN
Contact-Aided Navigation of Flexible Robotic Endoscope Using Deep Reinforcement Learning in Dynamic Stomach
Use contact as guidance, not simply as an obstacle.
1Department of Electronic Engineering, The Chinese University of Hong Kong · 2Department of Mechanical Engineering, Stanford University · IEEE ROBIO 2025
🏆T.J. Tarn Best Paper Award in Robotics — IEEE ROBIO 2025
ROBIO 2025
100%Static and dynamic success
1.6 mmAverage target error
85%Unseen disturbance success
01 / Overview
Why contact matters
The stomach wall can extend the robot’s reach.
A flexible endoscope cannot reach every gastric target through free-space bending and insertion alone. Contact with the stomach wall provides an external constraint that can support and redirect the endoscope.
CAN formulates this interaction as a reinforcement-learning problem. Force feedback becomes part of the policy observation, allowing the robot to exploit contact while navigating a moving, deformable stomach.
Dynamic stomach
A finite-element environment models deformation and physiological motion.
Unseen conditions
The trained policy is tested on new targets and stronger disturbances.
02 / Method
Simulation and learning
A force-aware policy trained in deformable anatomy.
The stomach and endoscope are modeled in the SOFA physics engine using finite elements and real-time contact detection. The simulation includes breathing, heartbeat, and other motion that changes the geometry during navigation.
A Proximal Policy Optimization agent receives the tip-to-target relationship, robot velocity, cable lengths, contact state, and force feedback. It learns a five-dimensional continuous action for cable-driven navigation.
03 / Project video
Paper presentation
Method and evaluation in 105 seconds.
The narrated video explains the contact-aided strategy, training environment, and navigation results.
04 / Results
Navigation performance
Stable control under deformation and disturbance.
Success is defined as reaching within 3 mm of the target. CAN reaches 100% success in static and dynamic stomach environments with an average error of 1.6 mm. In unseen cases with stronger external disturbances, it maintains an 85% success rate.
Target tolerance used to define a successful trial.
Continuous control learned through reinforcement learning.
Physics-based deformation and contact simulation.
05 / Citation
Publication
Contact-Aided Navigation.
IEEE ROBIO 2025, winner of the T.J. Tarn Best Paper Award in Robotics. DOI: 10.1109/ROBIO66223.2025.11376000.
@article{ng2025can,
title={Contact-Aided Navigation of Flexible Robotic
Endoscope Using Deep Reinforcement Learning
in Dynamic Stomach},
year={2025}
}