Dynamics & Control
Koopman operators, delay compensation, and data-driven adaptation for nonlinear systems.
01
Dynamics & Control
The classical feedback loop updated with modern learning-based tools for nonlinear, delayed, and data-rich systems.
Koopman Operators for Control
Koopman theory lifts nonlinear dynamics into a higher-dimensional space where the system evolves linearly. This enables the use of powerful linear control and estimation tools — such as LQR and Kalman filtering — on inherently nonlinear robots. Rather than wrestling directly with nonlinearities, the system learns a representation where prediction and control become linear algebra problems.
Observation Delay Compensation
Real-world sensors and networks introduce latency. If a humanoid or drone acts on stale observations, the Markov property is violated and standard RL theory breaks down. Recent work addresses this through augmented state representations and resampling techniques that reconstruct or compensate for missing and delayed information before it reaches the controller.
Data-Driven Adaptation
Modern dynamics and control is no longer purely model-first. Data continuously refines models — whether through online Koopman updates, delay estimators, or adaptive gain tuning. The flowing data streams connecting all modules in the research vision emphasize that learning and control now co-evolve in closed loop.
02
Agentic & System Modeling
Scaling from individual complex morphologies to multi-robot systems through learned simulators and local coordination.
FALCON
A dual-agent RL framework enabling humanoid robots to simultaneously locomote and manipulate heavy objects while adapting to interaction forces. Traditional approaches decouple walking from arm control, which fails when the robot must lean into a door or pull a cart. FALCON learns a unified force-aware policy.
CableRobotGraphSim
A Graph Neural Network simulator for cable-driven robots. These systems are notoriously hard to model because dynamics depend on tension, friction, and geometric configurations that are partially observable. CableRobotGraphSim uses rigid-body nodes and cable/contact edges to enable accurate prediction without full-state measurements.
MARL Locality
Scalable Multi-Agent RL where agents coordinate only within their local neighborhood. By factorizing value functions or policies over local graphs, teams of robots achieve near-global coordination without central bottlenecks — critical for swarm robotics and distributed autonomy.
03
Safety & Robustness
Wrapping learned components in rigorous statistical and control-theoretic safety certificates.
Conformal Coverage Guarantees
Conformal prediction turns point predictions into sets or regions with guaranteed coverage probabilities. If a learned model predicts a pedestrian's future position, conformal methods output a predicted region such that the true position lies inside with a user-specified probability — for example, 95%. This provides a rigorous statistical foundation for safe planning in uncertain dynamic environments.
Latent Control Barrier Functions
Control Barrier Functions define a "safe set" and ensure the system never leaves it. Rather than hand-designing these in the full state space, recent work learns CBFs in a low-dimensional latent space. These differentiable, learnable safety filters can be integrated directly into end-to-end training, making safety a native component of the learning pipeline rather than an afterthought.
04
Game Theory & Planning
When robots share spaces with humans or other robots, they must plan not just against nature, but against other decision-makers.
Markov Games & Coarse Correlated Equilibria
A Markov game generalizes a Markov Decision Process to multiple agents. Finding Nash equilibria is computationally hard, so researchers turn to Coarse Correlated Equilibria (CCE) — a more tractable solution concept where a coordinator can recommend strategies that no agent wants to unilaterally deviate from. This is critical for merging, intersection navigation, and collaborative lifting where agents must reason about each other's goals.
DyPNIPP: Dynamic Informative Path Planning
DyPNIPP addresses exploration in changing environments such as ocean monitoring and search-and-rescue. Rather than assuming a static map, it uses reinforcement learning to plan paths that maximize information gain while predicting how the environment itself evolves. As new data arrives, the robot updates its model of the world and recomputes its path — a continuous loop of replanning under uncertainty.
★
Cross-Domain Integration
The future of robot learning is not better isolated algorithms, but tightly integrated systems where dynamics, safety, learning, and strategic reasoning co-design.
The humanoid figure at the center of the research vision embodies the synthesis of all four themes. Each capability layer builds on the others:
| Capability Layer | What it Represents | Source Technique |
|---|---|---|
| Learned model with Koopman structure | The robot's internal dynamics model is linearized in a learned latent space, enabling efficient prediction and control. | Dynamics & Control |
| Models multi-jointed body with GNNs | Like CableRobotGraphSim, the robot represents its own body as a graph of links and joints for accurate simulation. | Agentic & System Modeling |
| Checks safety with conformal guarantees | Before executing a motion, it verifies that predicted trajectories stay within probabilistically safe regions. | Safety & Robustness |
| Plans based on opponent strategy | Using game-theoretic equilibrium selection, it anticipates how humans or other robots will react to its actions. | Game Theory & Planning |