My Dissertation Synopsis

I have defended my dissertation, and I wanted to share an easy to read summary of it. The full document is over 150 pages long, which is a lot to ask of anyone. This short version covers what I worked on, why it matters, and where I think the work goes next. I wrote it for a general engineering audience, so you do not need a background in robotics to follow along.

The Big Idea

The human brain does not represent the body in a single fixed way. It maintains a body schema, an implicit model that updates continuously as we move and act. This model is not uniform, since it devotes far more neural resolution to the hands and face than to the torso. It is also not static. It changes with context and can even expand to include a tool in the hand. This ability to meet each new task with a suitable representation is one of the quiet feats of biological intelligence.

Robots also depend on internal representations to reason about their bodies and surroundings. These representations determine how a robot estimates distance, detects collisions, and predicts the consequences of motion. Nearly every physical task depends on these calculations. A robot performs them constantly, so they also account for a large share of its computation and energy.

Yet in many systems, the representation is chosen once, often by hand, and held fixed for the whole task. The robot uses the same geometry whether it is moving through open space or squeezing through clutter, even though these situations place very different demands on it.

My PhD supports the following thesis. Adapting the geometric representation of a robot to its environment leads to better physical interaction, better use of resources, and better task completion.

Three Works, Each Building on the Last

I demonstrate this thesis through three first-author works. CAT-RRT was published at the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) in 2023. MorphIt was published in 2026 in the International Journal of Robotics Research (IJRR), one of the leading journals in the field. The third work treats the shape of the robot as something that can change while it runs, and it is under review at the IEEE International Conference on Robotics and Automation (ICRA).

Each work builds on the last. The first augments the representation for a single planner. The second automates the representation for any robot and task. The third adapts it online. The same ideas also shaped collaborative work on accurate particle-based simulation, whole-body tactile sensing (ICRA 2026), and safe control during contact (IFAC 2026).

Contribution 1 - Letting Robots Touch Things on Purpose

The first contribution asks whether a richer representation could expand the behaviors available to a single planner. Traditional motion planners treat the robot as one solid body and reject any path that produces contact anywhere on it. This is too strict for cluttered or shared spaces, where incidental contact may be harmless, unavoidable, or even useful.

I developed CAT-RRT, a contact-tolerant planner that represents the robot as individual links. Each link has its own cost function and its own limit for how much contact is acceptable. One link can stay in contact with an object while the rest of the arm moves freely.

When the goal itself required contact, two widely used planners failed in every trial while CAT-RRT always succeeded. In the most cluttered scenarios, it was compared with the only other method that always succeeded. CAT-RRT found paths in less than a third of the time, and those paths were up to three times shorter. A richer body representation turns contact from a planning failure into a controllable part of the interaction.

However, this representation was still designed by hand for one planner and one kind of task.

Contribution 2 - Generating the Right Shape Automatically

The second contribution removes that limitation. It asks whether robots could automatically generate the representation each task needs, rather than relying on hand-designed approximations.

I created MorphIt, a framework that approximates the shape of a robot with spheres, starting from its 3D model. Spheres are popular in robotics because the distance between two of them takes a single calculation. Existing sphere models, however, are placed by hand or generated by graphics methods that aim for visual similarity rather than useful robot behavior.

MorphIt uses gradient-based optimization with a loss that balances three goals. These are covering the volume of the robot, matching its surface, and keeping the spheres well organized relative to one another. By changing the weights, the same robot can be represented conservatively for collision avoidance, precisely for contact modeling, or efficiently for general planning.

MorphIt generated models up to two orders of magnitude faster than established methods, with half as many false collision alarms. In cluttered planning benchmarks it raised success rates from below 48 percent to 58 percent. In whole-body contact planning it improved convergence to the goal by about 53 percent. Because it plugs directly into existing pipelines and simulators, other laboratories have adopted it for their own robots.

Still, each MorphIt model is generated before deployment and stays fixed while the robot works.

Contribution 3 - Changing Shape While Working

The third contribution closes this gap by adapting the representation online. Real robots face changing demands within a single task, so a single fixed model is always a compromise.

Using MorphIt, I built a nested library of body models for each link, from coarse to fine. Every level comes with a guaranteed bound on how far it can deviate from the finest model. During operation, the robot picks the coarsest model that is still detailed enough to make the right decision. It bases this choice on obstacle proximity, motion speed, computational limits, and perceptual uncertainty.

The chosen model is shared by the planner and by a safety controller. The controller adjusts commands hundreds of times per second to keep the robot clear of obstacles and people.

Across motion planning, reactive control, and live human-robot interaction, adaptive morphology matched the collision accuracy of the finest fixed model. It did so at a planning cost close to that of coarse models, and it completed constrained tasks while keeping the whole body clear of obstacles.

Where This Goes Next

The most innovative part of this work is the shift it introduces to model-based robotics. It moves robots away from representations that are fixed before deployment and toward representations that adapt at runtime. A robot can then select among models, representations, and algorithms as task demands change.

In the short term, the next step is to extend these representations from avoiding contact to modeling it. Compliant sphere lattices, which I began developing with collaborators, represent contact as many small compliant spheres. They offer a fast middle ground between simple rigid-contact models and costly high-fidelity simulation. This could make it practical to generate the large interaction datasets that robot learning requires.

In the long term, the same principle of adaptation extends beyond geometry. It applies to perception, planning, and computation, so a robot can adjust how it senses, reasons, and acts as its tasks change. This will matter most for humanoids and other general-purpose robots now entering homes, hospitals, and warehouses. These robots face many different tasks on limited battery power. They cannot depend on engineers to hand-tune a new model for each one. Robots that direct their limited resources toward what each task demands will be able to operate reliably in the real world for long periods of time.

Wrapping Up

If there is one thing I hope you take away from this, it is that a robot does not need one perfect model of itself. It needs the right model for the moment. People do this without thinking, and I believe robots can learn to do it too.

Thank you for reading. If any of this sparks a question or an idea, I would love to hear from you.