Research Agenda 1.0 Beta

My research goal is to build robots that become better at physical interaction over time. I am especially interested in tasks where geometry alone is not enough: the robot must establish contact, regulate force, interpret partial physical feedback, recover from mistakes, and eventually turn its own experience into a better policy. The agenda is a learning loop: acquire physical priors from humans, ground them through robot contact, and improve them through the robot's own experience.

What I learned in the first year

At the beginning of my PhD, I thought tactile sensing was the center of my research. During the first year, work on human interaction data and force-aware manipulation changed how I see touch. It is an important instrument, but it is not the research question by itself. The larger question is how a robot acquires physical intelligence that remains useful when contact conditions, objects, embodiments, and tasks change.

This shift matters because a sensor-centered agenda can become a sequence of modality papers: add touch, improve a representation, fuse it with vision, and report a gain. A robot-centered agenda asks a harder question: what knowledge is missing from the policy, where can that knowledge come from, and how can the robot keep improving after deployment? Touch and force matter when they answer that question.

One research loop, viewed in three dimensions

Contact-rich manipulation, learning from humans, and continual learning are not three independent themes competing for equal attention. They play different roles in one causal program. Contact-rich manipulation is the problem domain. Human experience is the source of an initial physical prior. Continual learning is how that prior becomes reliable after deployment. The XYZ view below is therefore a diagnostic map, while the arrows are the research claim: human priors must be grounded in the robot's body, and failures during grounding must become new training data.

The problem: Contact-rich manipulation

Contact-rich manipulation defines the class of problems I want to solve. Tool use, insertion, pivoting, surface following, in-hand adjustment, and articulated-object interaction cannot be reduced to a desired end-effector trajectory. Two executions with nearly identical visual trajectories can produce different outcomes because of friction, object mass, contact location, compliance, or force direction.

Physical grounding

I want to move manipulation from a trajectory-centric formulation toward an interaction-centric one. A policy should reason about where contact is established, how the contact mode evolves, what forces are compatible with task progress, and when a mismatch requires recovery. This motivates my current work on a predictive-and-reactive force policy: a slower component generates task-level motion and an expected interaction process, while a faster feedback component uses force, touch, and proprioceptive history to adjust local motion or compliance.

Force policy is therefore a technical direction, not the final identity of the agenda. Its value should be measured by whether it enables manipulation under uncertainty that position control, fixed impedance, or a purely visual policy cannot handle reliably.

The source: Human physical priors

Human experience is the most scalable source of complex manipulation structure we currently have. People naturally demonstrate tool use, dexterous contact, and long-horizon procedures. But a human demonstration is not a robot action label. Human hands and robot hands differ in morphology, actuation, compliance, sensing, and reachable contact configurations. Copying motion frame by frame can transfer the wrong thing.

My goal is to extract what is physically transferable: task intent, object progress, functional contact regions, interaction phases, plausible pressure patterns, and motion priors. The robot should inherit this structure while remaining free to discover an embodiment-specific solution.

OpenTouch is an important foundation for this direction. By synchronizing egocentric video, full-hand touch, and hand pose, it connects visual semantics and motion with actual physical contact. In the longer term, data of this kind could provide contact priors for human videos that contain rich task information but no direct tactile measurement. The central challenge is cross-embodiment grounding: transferring interaction structure without assuming a shared action space.

The learning mechanism: Continual improvement

A policy learned from demonstrations is a starting point, not a finished robot. Deployment introduces new materials, objects, tools, mounting errors, sensor drift, and contact geometries. A robot should use these interactions—especially its failures and recoveries—to improve rather than remain frozen at the moment training ends.

I use continual learning in three time scales. The first is adaptation within one execution: recent force, tactile, and proprioceptive history should change the next action immediately. The second is improvement across attempts: failures, corrections, and successful recoveries should update the policy with limited additional supervision. The third is long-term accumulation across tasks and embodiments: new skills should reuse contact primitives and interaction structure without destroying capabilities the robot already has.

This formulation makes continual learning a physical systems problem. Data is expensive and actively generated by the current policy; some failures are unsafe; feedback is partial and noisy; and a local disturbance should not always trigger permanent model change. The robot must decide which experiences are worth retaining, when to update, and how to improve under new conditions without forgetting old ones.

How my current work fits

I keep only the three projects that define the causal spine of this PhD agenda:

• OpenTouch — Human Priors. Capture real-world full-hand contact together with vision and pose, so the robot can begin with physical interaction structure rather than learn every contact pattern from scratch.
• Force Policy — Physical Grounding. Convert an abstract contact prior into embodiment-specific action by predicting task-level interaction and reacting to force at a faster control time scale.
• Contact DAgger — Continual Improvement. Let the robot visit its own contact states, request targeted human corrections when interaction departs from the intended regime, and aggregate those recoveries into the next policy. It reframes DAgger's state-distribution shift as a contact-distribution shift.

The 2026 agenda

In the near term, I want to establish when learned force-aware policies provide a real advantage. The comparison should be against strong position policies, fixed impedance control, and hand-designed hybrid force control—not only against a weak baseline. Evaluation should vary friction, object mass, contact geometry, perturbations, force bias, and sensor noise. The important outcomes are task success, peak force, recovery time, and generalization to unseen contact conditions.

The next step is to connect OpenTouch-style human interaction data to robot policy initialization. I want to study joint visual–pose–touch pretraining, functional contact representations, interaction phases, and contact retargeting. The test is simple to state: can human contact data reduce the amount of real robot data required while improving generalization to unseen objects, tools, and physical conditions?

The longer-term system closes the loop. Human experience provides an initial physical prior. The robot grounds that prior in its own embodiment, executes a contact-rich task, diagnoses mismatches through physical feedback and task outcomes, stores useful recoveries, and improves the next execution. Each cycle moves the robot from imitating available experience toward forming its own.

What would count as progress?

A coherent agenda needs shared evaluation, not one metric for each paper. I want to measure whether a system can start from less robot data, operate under wider physical variation, recover from perturbations, and improve across attempts without losing previous abilities. These criteria test the entire loop.

1. Does human experience reduce robot data requirements rather than merely add another pretraining loss?
2. Does contact sensing change the action when vision and geometry are ambiguous?
3. Does adaptation improve recovery under controlled physical perturbations?
4. Does learning from new interaction preserve previously acquired skills?
5. Can the same interaction representation transfer across tasks, tools, sensors, or embodiments?

These questions also prevent the agenda from becoming “tactile for tactile’s sake.” A new sensor, representation, model, or controller is useful only when it improves the larger loop from human prior to grounded interaction to continual improvement.

The one-sentence version

My research develops continually improving robots for contact-rich manipulation. I study how robots can learn transferable physical interaction priors from human experience, ground them in embodiment-specific manipulation policies, and continually improve through real-world interaction.

This is the research identity I want to develop after year one: not a collection of tactile projects, but a connected program for robots that learn from people, understand contact, and become better through experience.