Tim Seyde

I am a final-year PhD student at MIT CSAIL advised by Professor Daniela Rus. My research focus is on sample-efficient learning for robot control based on concepts from model-based exploration, architecture simplification, and compositionality. Before joining MIT, I obtained a BSc in Mechanical Engineering and an MSc in Robotics and Control from ETH Zurich under supervision of Professor Marco Hutter. Throughout my studies, I had the opportunity of various research stays and internships in legged robotics as well as learning control, and was fortunate to learn from many brilliant mentors.

.

I'm interested in reinforcement learning, motion planning, and control. I'm particularly excited about application in (legged) robotics, leveraging learning-based approaches to unlock capabilities that prove challenging for traditional methods.

Summary: we combine the benefits of coarse exploration during learning and smooth control at convergence by growing action spaces via decoupled Q-learning.

Summary: we leverage a hierarchical policy design for multi-team dynamic flight control, enabling high-level strategic coordination via distributional decoupled RL.

Summary: we present Gigastep, a fully vectorized JAX-based MARL environment that features high-dimensional 3D dynamics, heterogeneous agent types, stochasticity, and partial observation.

Summary: we deploy a hierarchical RL agent in multi-team racing scenarios on scale-car hardware, combining decoupled SARSA for team-centric strategic reasoning with continuous PPO for ego-centric low-level decision making.

Summary: we show that DQN combined with value decomposition and bang-bang action space discretization yields performance competitive with recent model-free and model-based actor critic algorithms when training from features or raw pixels.

Summary: we analyze how decision trees based on logic programs extracted from a compact bio-inspired model architecture can help interpretable decision making.

Summary: we investigate a parallel autonomy system for flight based on attention map mismatches with a bio-inspired policy architecture.

Summary: we investigate replay memory interpolation as a data augmentation technique for improving data efficiency of reinforcement learning agents.

Summary: we propose a suite of challenging problems to test agents in autononmous flight scenarios towards guardian autonomy systems.

Summary: we show how a bio-inspired model architecture yields compact policies for solving autonomous flight scenarios.

Summary: we show that several recent actor critic algorithms yield competitive performance when only considering bang-bang policy heads and discuss implications for agent and benchmark design.

Summary: we leverage a hierarchical model over diverse low-level policy architectures to transfer the burden of hyperparameter selection from the engineer to the agent.

Summary: we leverage a latent model ensemble to compute an upper confindence bound objective over predicted returns to guide exploration in continuous control from pixels.

Summary: we leverage self-play in latent space with a world model that estimates opponent behavior to generate competitive racing maneuvers.

Summary: we learn a value function that represents an upper confidence bound over expected returns to guide exploration in continuous control from features.

Summary: we solve a mixed-integer gait planning problem for a single-legged hopper by learning footstep selection with a Gaussian process, and using this to constrain a low-level trajectory planner.

Summary: we augment a capture-point based walking controller to account for swing-leg angular momentum during reference trajectory planning and show improved locomotion capabilities with an Atlas humanoid robot.

Summary: we implement a bio-inspired modular posture controller and compare to model-based control on disturbance compensations tasks with the TORO humanoid.

The template is taken from here. Thanks so much!