[ICON Fall26 Seminar] Johannes Lachner (Purdue) #robotics #control #physicalai
Johannes Lachner, assistant professor of ECE at Purdue with a Mechanical Engineering appointment, presents research on physics-grounded AI for stable, safe robot interaction, with examples from rehabilitation, industry, and prosthetics.
In this talk
- 3:02
What robot control can learn from people
Lachner compares robots’ fast motion generation with people’s ability to adapt during physical contact. He argues that internal predictions and motor primitives, such as bell-shaped reaching motions and walking oscillations, can help robots manage interaction.
- 7:30
Geometry, energy, and safe interaction
Lachner describes using differential geometry to model robot behavior in ways that do not depend on the chosen coordinates. He says this supports stable impedance control and energy-based safety: limiting stored spring energy can let a robot adjust its compliance near a person and continue working after contact.
- 22:46
Combining learned motion and impedance
Lachner presents two approaches to robot primitives. Contraction theory lets learned discrete and oscillatory motions be combined while preserving stability, while a diffusion model learns impedance changes from human force and motion data. In a rehabilitation example, the robot reduced assistance when a participant contributed more and increased it when they contributed less.
- 32:29
Rehabilitation and industrial examples
Lachner describes collecting therapist motion and force data with cameras and gloves to train a model for guiding patients. He also reports that a model trained on human experiments adapted robot impedance during peg insertion, including with shapes it had not encountered in training. He says the approach still needs further study, including how to generate motion as well as stiffness.
- 37:53
Safety layers and future directions
Lachner proposes applying the methods to prosthetics, with diffusion models conditioned on EMG or EEG signals. In the discussion, he explains that an energy check limits unsafe impedance increases and can fall back to gravity compensation. He sees motion generation and impedance generation as complementary layers, while noting that end-to-end methods would need to provide comparable safety guarantees.
Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.