On 13 February 2026, the MIT Computer Science and Artificial Intelligence Laboratory (CSAIL) hosted a presentation by Dr. Gabriel Margolis titled “Learning Intelligent Contact for Dynamic Robots.” The talk, introduced by Dr. Pulkit Agrawal, detailed Dr. Gabriel Margolis’s research into force-aware robot learning, moving beyond traditional position-based control toward systems that perceive and manipulate the physical language of force.

The Evolution of Force in Robotics

Dr. Gabriel Margolis opened his presentation by grounding modern robotics in historical context, citing early MIT projects like the Raibert Hopper from the 1980s. He explained that while wheeled robots move by spinning actuators, underactuated legged robots must manipulate the magnitude and direction of forces applied through their limbs to achieve locomotion. Drawing a distinction between force-unaware industrial robots—which often cause damage when encountering unexpected obstacles—and humans, who excel at forceful manipulation, Dr. Gabriel Margolis argued that force first robot learning is essential for making dynamic machines safer and more capable.

Sim-to-Real and Data-Driven Control

A significant portion of the talk focused on the shift toward data-driven learning in simulation. Dr. Gabriel Margolis described his early work with the MIT mini cheetah, where he helped establish a reinforcement learning stack that enabled agile motions and fast running. He noted that unlike analytically designed controllers, these policies are learned through experience in simulators capable of running thousands of robots in parallel. This paradigm was further demonstrated through a “dribble bot” project, where a quadruped learned to manipulate a soccer ball across diverse terrains, including snow, by balancing the needs of locomotion with the forces required for ball control.

Learning Force Control via Virtual Fields

To address whether data-driven systems can explicitly sense force without dedicated sensors, Dr. Gabriel Margolis introduced a method involving virtual force fields. By spawning a virtual set point that attracts the robot’s gripper with a specific resistance, the system creates a soft contact model during training. This allows the robot to learn to feel its way through resistance based on feedback from its joints. Dr. Gabriel Margolis showcased how this enabled a quadruped manipulator to remain compliant when touched by humans or to compensate for the weight of a heavy object while remaining responsive to external guidance.

Active Sensing and Informative Probing

Dr. Gabriel Margolis also detailed research into active sensing policies, which optimize contact to gain information about the environment. In scenarios where a robot must choose a path based on friction, it may not be able to determine surface properties simply by walking. By rewarding the estimation of physical properties during training, Dr. Gabriel Margolis showed that robots can develop emergent probing behaviors, such as scuffing a foot to generate shear forces that reveal the friction of the terrain. This information can then be used to train vision models for future traversal.

Humanoid Compliance and Safety

The final pillar of the presentation addressed humanoid robots through a project called Soft Mimic. Dr. Gabriel Margolis explained the challenge of training humanoids to imitate human motions while maintaining safety. By reframing compliant whole-body control as an augmented motion-imitation problem, he demonstrated that a humanoid could follow a reference motion—such as reaching for a box—while remaining force-aware. If the robot encounters an obstacle, such as a misplaced box or a human hand, it gracefully yields rather than attempting to smash through it, a behavior that emerges from the generalized compliant objective rather than explicit programming for every possible obstacle.

Future Directions in Force-Aware Learning

In his concluding remarks, Dr. Gabriel Margolis reflected on the limitations of current models, such as the reliance on manual tuning for stiffness levels and the need for broader information-gathering strategies like world models. He was joined by committee members Dr. Daniela Rus and Dr. Marco Hutter, as well as Dr. Pulkit Agrawal, who praised Dr. Gabriel Margolis for his role in establishing the lab’s legged locomotion capabilities and his mentorship of fellow students throughout his doctoral studies.

The Computer Science and Artificial Intelligence Laboratory (CSAIL) is MIT’s premier research center for computer science and artificial intelligence, driving advances in robotics, machine learning, and cybersecurity. By uniting experts across disciplines and partnering with industry, CSAIL cultivates innovations that solve real-world problems and train the next generation of technology leaders.

The Conf is a platform that reports on scholarly conferences, symposia, roundtables, book talks, and other academic events. It is managed by a group of students from leading American and European universities and is published by Alma Mater Europaea University, Location Vienna.