Anthony Simeonov

I am a Ph.D. student in the EECS department at MIT, advised by Professor Alberto Rodriguez and Professor Pulkit Agrawal. I'm interested in robot manipulation and physical interaction. I am fortunate to be supported by the NSF GRFP.

I received my S.M. in EECS from MIT and my B.S. in Mechanical Engineering from UC San Diego, where I worked with Professor Michael Yip. I have also spent time as a research intern at Disney Research Los Angeles and NVIDIA's Seattle Robotics Lab

Email  /  Google Scholar  /  LinkedIn

I'm currently interested in machine learning for perception-driven planning and control, applied to problems in robot manipulation and physical interaction. The goal is to equip robots with reusable manipulation skills, through a combination of effective perceptual representations and reactive behaviors. Previously, I have studied the design, modeling, and control of inherently compliant artificial muscle actuators.

A BC + RL pipeline for precise manipulation tasks like assembly. Fine-tuning BC-trained policies with action chunking and diffusion de-noising requires specialized RL methods. We simplify such RL fine-tuning with simple residual policies trained with PPO.

Learning robust manipulation policies by combining real-world imitation learning and reinforcement learning in simulation. RL in sim uses digital twin assets that reflect the geometry, appearance, and kinematics of the target deployment scene.

A pipeline for learning image-based policies for precise, long-horizon assembly tasks from a small number of demonstrations by combining expressive policy architectures and various techniques for dataset expansion

An LLM-based task planner that can learn new skills opens doors for continual learning.

Performing 6-DoF relational rearrangement between objects and scenes. Iterative pose de-noising via diffusion helps handle multi-modality and local scene cropping improves generalization.

Local point cloud encoders help 3D neural descriptor fields generalize to more diverse global shapes and improve robustness to occlusions

Applying 3D neural descriptor fields in a pairwise fashion, to enable relational rearrangement with pairs of unseen objects in arbitrary poses.

NeRF enables synthesizing orthographic views from virtual camera poses, allowing search for 6-DoF placing actions represented by pixels in the rendered image (positions) and the camera pose used to generate the image (orientations)

A novel representation that models objects as 3D neural fields of descriptors. We apply this representation to enable pick-and-place on unseen objects in out-of-distribution poses from a small handful of demonstrations.

A framework for solving long-horizon planning problems involving manipulation of rigid objects that operates directly from a point-cloud observation

By encoding raw observations of obstacle geometry into a latent representation, we can distill the capabilities of an expensive planning algorithm into a neural network, allowing significant speedups in finding near-optimal collision-free paths.

This paper is subsumed by our journal paper.

Empirical study of the performance tradeoffs between different bundling techniques for thermally-driven, compliant artificial muscle actuators.

We present a two DOF robot that is launched from a gravity-driven pendulum and executes a variety of mid-air stunts. The robot uses a combination of onboard sensors to perform state estimation and inform actuator timing.

This research is part of the basis of Imagineering's Stuntronics project.

A new modeling scheme for capturing the coupled hysteresis between tension, strain, and input, which is used for improved control of artificial muscle actuators.

We propose three new models for characterizing the hysteresis present in super-coiled polymer artificial muscles, and demonstrate their use for accurate open-loop control.

template