MCP-EP | Intelligent Robotic Systems Lab (IRSL)

Proprioceptive sensing
Particle filtering

A multi-contact particle filter combines preprocessed robot meshes, exploration particles, and a base force/torque sensor to estimate contact locations and forces without tactile skin.

IEEE ICRA · 2023

Simultaneous Multi-Contact Localization and Force Identification

Locating Contacts and Identifying Forces

MCP-EP estimates contact positions and forces from joint torque sensors and a force/torque sensor at the robot base. The base measurement adds six sensing degrees of freedom, helping distinguish contacts that joint torque measurements alone cannot identify, particularly on links near the base. The method supports several simultaneous contacts without a tactile skin.

MCP-EP estimates contact locations and forces from joint torque and base force/torque measurements — Multi-Contact Particle Filter with Exploration Particles.

Fast Particle Updates on Robot Meshes

The robot mesh is preprocessed into nearly uniform faces, and geodesic distances between faces on each link are computed offline. A particle can then move over the surface through a lookup and face-index update. For a proposed contact location, a convex quadratic program finds forces consistent with the sensor measurements and a polyhedral friction cone. Its residual determines the particle weight.

Robot mesh preprocessing removes non-contact regions and creates nearly uniform surface faces — Mesh preprocessing makes particle positions and surface updates inexpensive.

Precomputed geodesic distances between faces on a robot link — Distances within each link are stored before online estimation.

Exploration Across Links

Distances between different links change as the robot moves and cannot use the same fixed lookup. Exploration particles search adjacent links, allowing the estimate to move away from an incorrect link. Particle sets are added or removed when the current contact hypotheses no longer explain the sensor measurements, or when a set becomes unnecessary.

Particle motion using precomputed surface geometry.

Exploration particles search neighboring links.

Simulation Validation

Each contact condition was evaluated in 10,000 randomized trials using a seven-degree-of-freedom Franka Emika Panda in MuJoCo. A 20 N force was applied at each sampled contact, with at most one contact per link. Success means localization error no greater than 2.25 cm. The reported RMSE values include failed trials.

Simulation results (Table I); these are not real-world success rates.
Contacts	Success	Position RMSE	Force RMSE	Runtime
1	99.96%	0.16 cm	0.01 N	0.45 ms
2	96.70%	1.08 cm	2.00 N	1.66 ms
3	81.63%	3.50 cm	4.63 N	4.50 ms

Supplementary Real-World Demonstrations

The videos show contact estimation on a Kinova Gen3. These demonstrations are separate from the simulation statistics reported above.

Supplementary triple-contact estimation on the Kinova Gen3.

MCP-EP supplies the contact estimator in Multi-Contact Feedback MPC and serves as a comparison method for ContactFlowFilter.

Representative Publications

Proprioceptive Sensor-Based Simultaneous Multi-Contact Point Localization and Force Identification for Robotic Arms

Seo Wook Han and Min Jun Kim

IEEE International Conference on Robotics and Automation (ICRA), 2023

Paper

Bib

@inproceedings{han2023proprioceptive,
  title={Proprioceptive Sensor-Based Simultaneous Multi-Contact Point Localization and Force Identification for Robotic Arms},
  author={Han, Seo Wook and Kim, Min Jun},
  booktitle={IEEE International Conference on Robotics and Automation (ICRA)},
  pages={12099--12105},
  year={2023}
}

Youtube Project

Figures, demonstrations, and research details: author's project page.