• bioinspiration
• elephant trunk
• learning from demonstration
• robot control
• soft robot
• trajectory planner

Soft robots benefit from high deformability, enabling compliance and adaptability for interaction but posing control challenges. To tackle this, researchers draw inspiration from animals, such as elephants and their trunks, that exhibit similar features to these robots. This work aims to develop a bio-inspired trajectory planner for a soft robotic arm by studying how an elephant trunk performs reaching-for-grasping of target objects.
Elephant trunk is recorded while performing goal-directed reaching-for-grasping tasks using RGB-D cameras, which provide depth and colour information about the surroundings.
These recordings are post-processed to extract meaningful data about the trunk motion, with steps including scene reconstruction, trunk segmentation, and skeleton extraction. Tip trajectory data is then analyzed in terms of kinematic invariances, specifically Power Laws and the Minimum Jerk Model. These analyses provide insight into how elephants execute smooth and coordinated reaching movements, serving as a foundation for the planner design.
Biological data are then fed to a Learning-from-Demonstration framework. Trajectories are modelled using a combination of Gaussian Mixture Models (GMM), Gaussian Mixture Regression (GMR), and Dynamical Movement Primitives (DMP). GMM is used to generalize across multiple demonstrations of trunk movements towards the same target, capturing the natural variability in the elephant reaching strategies. GMR is then applied to extract an ideal, averaged trajectory from these demonstrations. DMPs, encoding movement as dynamic systems, are used to generate trajectories for the robot, ensuring that it can replicate the elephant behaviour in reaching tasks.
An adaptive system is integrated into the planner to modulate the generation of trajectories for new targets. This approach allows the generation of unobserved, smooth trajectories that share the characteristics of the biological movements, even for unseen targets.
The results demonstrate that the bio-inspired approach can successfully replicate the smooth, adaptive reaching movements observed in elephants. The integration of GMM, GMR, and DMPs, along with the adaptive system, enables the generation of versatile trajectories that maintain key features of the biological data while adapting to new task constraints.
This work offers overall a novel approach to motion planning that bridges biological insights with soft robotic control methodologies.