MIT Robotics - Michael Posa - Contact-rich robotics: learning, impact-invariant control, and tactile
Michael Posa, an assistant professor at the University of Pennsylvania and director of its Dynamic Autonomy and Intelligent Robotics Group, presents recent work on contact-aware control and learning at an MIT Robotics seminar in Cambridge.
In this talk
- 2:54
Why contact is hard to plan
Posa contrasts robots’ carefully choreographed parkour and household actions with the adjustments people make in unfamiliar environments. Contact makes planning difficult because surfaces, shapes, and internal dynamics vary, while impact forces can change abruptly and plans often need correction.
- 7:11
Control through impacts and touch
Posa describes a controller for Cassie that projects out velocity directions most affected by impact, reducing disruptive feedback during landings. He also presents tactile feedback methods that use measured force without explicitly detecting contact modes, including a car-shaped robot that stays stable against soft walls. A real-time contact-aware planner using ADMM handles a horizon of about 10 steps and three contacts at roughly 20 Hz.
- 20:41
Designing simpler robot models
To make real-time planning manageable, Posa’s group optimizes low-dimensional models for a set of tasks rather than choosing a model by hand. In simulation, adding model capacity reduced expected task cost by about 30 to 40 percent. The approach combines planning around 1 Hz with faster whole-body control, though the results described are still in simulation.
- 26:27
Learning contact from motion data
Posa argues that directly fitting a neural network to motion can miss brief, consequential events such as impacts, especially in stiff systems. His group instead learns contact geometry and related quantities that feed a physics solver, using observed motion to guide candidate contact forces during training. On cube experiments, structured models learned shape and friction from roughly five to ten throws, while a direct network struggled with orientation and penetration.
- 43:24
Open problems in dexterous tasks
Posa says standard learning and hybrid control methods can impose assumptions that do not fit contact-rich systems. He sees task-guided model learning as a path toward dexterous manipulation, where a planner needs compact models that still capture useful contact choices. The group is also exploring how to learn smooth motion and contact effects together.
Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.