Robotics in Europe: look back on ICCMA 2025

When one is going to Paris for a work trip, their mental images of the destination might vary, and the outcomes depend on the sort of work the individual does: someone involved with museums or history may be immersed in scenes much like the left photo; but someone engaged in a less posh line of work will end up with scenery resembling that on the right.

The author being one of the second camp, he didn't bemoan unvisited palaces and cafés, but instead took his dive into the newest robotics research presented at ICCMA'25 in Velizy. The author's industriousness doesn't stretch enough to review the whole of the event, but the projects which stuck with him are covered in what follows.

1. Drone swarm control for wild fire fighting

3D CFD Navier-Stokes equations are used to model controlled fire spreading via the reaction-diffusion model. Each chemical involved in the burning is modeled as a separate term in these equations. Certain LFMC quantity (Live Fuel Moisture Content) is derived using the tools of image processing from drone imagery captured with infra-red cameras. The simulation is done with historical wind and temperature data, manually injected the controlled fire origin and the natural fire barriers. Control is derived as PID using a potential field setup, to repeal fire from the safe region (e.g. a settlement).

2. Loader crane control

The automation for a knuckle boom crane was created, which employs a stereo camera with onboard VPU compute and IMU together with an industrial controller (Beckhoff) that achieves material handling. Camera's compute unit runs YOLO NN, that registers the manipuland and the goal region for every task. The crane is controlled with Inverse Kinematics PID. Gains for the control and hydraulics' dead-band offsets were established experimentally.

3. Underwater Gaussian Splats

NeRF is volumetric rendering method which is on the forefront of 3D reconstruction in terms of visual quality. However its training times limit feasibility of many applications. Gaussian Splats is a competing approach that offers comparable quality with training times several orders of magnitude shorter. However, the explicit representation inherent to GS makes it challenging to use it for underwater scenes, where semi-transparent bodies of water need to be represented.
Proposed method uses a combination of the two approaches, where GS are culled (with use of a 3D reconstruction pipeline) in the volume's regions corresponding to the semi-transparent media.

4. Safety Controller to augment Learned Robot Control Policies

Several approaches to strengthen control policies are investigated, in application to imitation- and reinforcement- learned models. The directions of policies' improvement are robustness and safety, where the former relates to the sim-to-real gap and success rate, and the latter ensures respecting the hardware's limits and reachability. The tasks selected for this experiment are peg-in-hole insertion and cube pushing.
The investigated approaches for imbuing safety onto learned policies are the training-time adjustments (e.g. domain randomization), including the kinematic limits violation penalty terms into the reward function, and clipping safety controller (with and without its use during train-time, relates to PLAI).

5. Trajectory optimization for energy efficiency

Minimize the total power used by a redundant robot for a specified trajectory tracking goal. The goal gets formulated in terms of control points in robot's inner coordinates. The torques are computed via inverse dynamics and control points via inverse kinematics. The approach is reported to obtain 40%+ energy-efficiency gains on a Franka Panda manipulator.

Closing Remarks

Besides the above 5, the author's own contribution to ICCMA is presented on the blog's frontpage. Overall, this conference was stimulating and impressive, and much has been left out of this post.

References

  1. Rafal Krzysiak, Andrew Amavizca, Jesse Ramirez, Derek Hollenbeck, Sachin Giri, YangQuan Chen, "Modeling and Control of a Prescribed Fire with UAVs as Sensors and Actuators"
  2. Lars Østerholt, Are Gundersen, Jing Zhou, Daniel Hagen, "Real-Time Markerless Vision-Based Control of Knuckle Boom Cranes via Embedded AI"
  3. Anurag Dalal, Daniel Hagen, Kjell G. Robbersmyr, Kristian Muri Knausgård "ScatteringSplatting: Capturing Minute Details Using Deep Learning-Aided Gaussian Splatting"
  4. Konstantin Wrede, Sebastian Zarnack, Yibo Di, Julius Neumann, Martin Dehmei, Ron Martin, Dirk Mayer, Peter Schneider, "Towards a Workflow for Safe Simulation-to-Reality Transfer of Robot Control Policies"
  5. Giuliano Fabris, Lorenzo Scalera, Alessandro Gasparetto "Experimental validation of energy-efficient optimal trajectories for redundant robotic systems"

The images in 1,3,4 are courtesy of respective researchers.