Zhaoting Li

PhD research at TU Delft, published at RSS 2026 [paper] [code]

Proposed Set-Supervised Diffusion Policy (SDP), which trains action-chunking diffusion policies from human corrections by pairing the robot’s undesired action-chunks with the human’s corrective action-chunks.
Used these pairs to construct sets of desired action-chunks and trained the diffusion policy to imitate samples from these sets, instead of only imitating positive actions.
Outperformed standard Diffusion Policy on real-robot Insert-T and round-table assembly tasks, with the largest gains on harder initial states and better robustness to noisy data.

PhD research at TU Delft [arXiv] [code]

Showed that behavior cloning, which treats each human action as an exact label, is brittle when the labels are imperfect, especially for expressive models such as energy-based models.
Proposed CLIC, which uses human corrections to build and refine sets of desired actions and trains the policy to place probability mass over these sets, supporting both absolute and relative corrections and multi-modal behaviors.
Demonstrated in simulation and real-robot experiments (Insert-T, water pouring) that CLIC is competitive with state-of-the-art methods on accurate data and substantially more robust to noisy, relative, and partial feedback.

2022.10 - 2023.06

MSc thesis in Computational Robotics Lab (CRL), ETH Zurich; RSS 2023 workshop paper [paper] [video]

Implemented a contact detection algorithm that estimates contact locations and forces from joint torque sensor data.
Extended the contact particle filter method to output multiple possible solutions, generalizing the state-of-the-art method to multiple-contact cases.
Implemented a contact-aware control framework that keeps interaction forces over the robot’s whole surface below a safety threshold, for both soft and rigid contacts.
Extended the BiTRRT sampling-based planner to be contact-aware, and validated the framework on a Franka robot in a setup inspired by the Amazon stowing task.

2022.03 - 2022.08

Semester project in Robotic Systems Lab (RSL), ETH Zurich [Report]

Proposed a learning-based force filtering method that uses the filtered-and-amplified robot-side forces as the ground truth, to reduce the noise introduced by the trotting action.
Designed an efficient transformer-based neural network, collected data during collaboration between humans and ANYmal with an arm, and trained the network in a supervised learning framework.
Tested the effectiveness of the proposed force filtering method with the admittance control framework.

2021.10 - 2022.02

Based on the robot exploration project

Developed the simulation environment for multi-robot exploration.
Proposed the adaptive-QMIX method, which adapts to a varying total number of robots.
Applied the proposed algorithm to the multi-robot exploration problem.

2022.04 - 2022.07

Course project of Deep Learning for Autonomous Driving, ETH Zurich

Trained neural networks based on three types of multi-task learning frameworks, and compared the effects of different hyperparameters and structures.
Proposed an adaptive weight tuning method that dynamically adjusts the weight of each task’s loss during training.
Trained the baseline proposal refinement network, and proposed several methods to improve its performance.

2020.10 - 2021.07

Research work at SUSTech, published at RA-L [paper].

Proposed 4D point-clouds-like information, which consists of 2D points’ location information, and the corresponding 1D frontier and 1D distance information.
Designed the corresponding training framework based on the deep Q-Learning method and modified it to adapt to the variable action space.
Demonstrated the performance of the proposed method on a wide variety of unseen environments, including maps much larger than those in the training set.

2020.07 - 2020.10

Research work at SUSTech, published at ICRA 2021 [paper].

Designed a novel recurrent generative model to generate efficient heuristics for robot path planning.
Incrementally constructed the heuristic through the feedback of historical information.
Combined the generated heuristic with the RRT* algorithm to guide it to find both initial and optimal solutions faster and more efficiently.

2019.07 - 2020.01

Research work at UC Berkeley, published at IFAC 2020 [paper].

Applied the discrete elastic-band-based motion planning method (EB planner) to generate a piecewise-linear collision-free path with dynamic programming.
Employed a pure pursuit controller to smooth this path, and used a spatial and speed sampling method together with a cascaded ranking method to optimize the trajectory with many hierarchical features.
Applied a non-conservatively defensive strategy to avoid overreacting to threats with low probability.
Adaptively adjusted the sampling resolution based on the environment and the objectives of the ego vehicle.

2019.02 - 2019.06

Research work at HIT, published at Neurocomputing 2020 [paper].

Designed an automatic and low-cost robot scanning system consisting of a Kinect camera, a UR10 robot and a line laser scanner, and calibrated the coordinate systems.
Proposed an online correction method based on follow-up control and scanned data to optimize the pose of the laser scanner.
Proposed a novel path planning method for laser scanning based on least-squares fitting and online correction, validated in many use cases under various working conditions.