Abulikemu Abuduweili

I am a PhD candidate in the Department of Electrical and Computer Engineering and Robotics Institute, at Carnegie Mellon University. I am fortunate to be advised by Prof. Changliu Liu, in the Robotics Institute, School of Computer Science. Before joining CMU, I spent an enriching year as a research assistant with Prof. Humphrey Shi. I hold a Bachelor's and a Master's degree from the School of Electronics Engineering and Computer Science at Peking University. My research focuses on Machine Learning, Robotics, Embodied AI, and Computer Vision.

I gained industry experience through internships at Microsoft Research Redmond (with Dr. Lekan Molu and Dr. Naoki Wake), Toyota Research Institute (with Dr. Frank Permenter and Dr. Chenyang Yuan), Baidu Research (under Prof. Dejing Dou), and Microsoft Research Asia (with Dr. Qingwei Lin). Feel free to reach out for collaborations or to discuss research.

I am actively seeking a full-time, research-oriented position. If my expertise aligns with your needs, please feel free to reach out.

Email  /  GitHub  /  Google Scholar  /  CV  /  LinkedIn

Research

I'm broadly interested in the intersection of robotics and machine learning, with a focus on areas such as continual learning, robot learning, human-robot collaboration, optimization, data-efficient learning, time-series prediction, test-time adaptation, visual perception, and generative models. Below is a list of my research works (* denotes equal contribution), with some papers highlighted. For the full list of publications, please visit my Google Scholar.

Enhancing Sample Generation of Diffusion Models using Noise Level Correction

This paper proposes a novel method to enhance diffusion model sample generation by aligning noise level estimates with the true distance to the underlying manifold. A noise level correction network refines estimates during denoising, compatible with any scheduler (e.g., DDIM), and extends to image restoration tasks like inpainting, deblurring, super-resolution, colorization, and compressed sensing.

Continual Learning and Lifting of Koopman Dynamics for Linear Control of Legged Robots

This paper presents a continual learning algorithm for iteratively refining Koopman dynamics in high-dimensional legged robots. The method ensures monotonic convergence of linear approximation error and achieves high control performance on Unitree and ANYmal robots across diverse terrains with linear MPC controllers.

Robustifying Long-term Human-Robot Collaboration through a Hierarchical and Multimodal Framework

This paper introduces a multimodal hierarchical framework for efficient long-term human-robot collaboration, integrating visual observations and speech commands for intuitive interactions. Deployed on the KINOVA GEN3 robot, the framework is validated through extensive real-world user studies.

Revisiting the Initial Steps in Adaptive Gradient Descent Optimization

This work identifies the standard initialization of Adam’s second-order moment estimation as a key factor limiting generalization. It proposes effective solutions, including non-zero initialization with data-driven or random strategies, to improve performance.

KOROL: Learning Visualizable Object Feature with Koopman Operator Rollout for Manipulation

This paper proposes a method to construct a Koopman operator using object features extracted from images, enabling precise propagation of robot trajectories. By embedding scene information into learned features, the approach achieves high task success rates and extends to real-world manipulation without ground-truth object states.

Estimating Neural Network Robustness via Lipschitz Constant and Architecture Sensitivity

This paper examines neural network robustness in perception systems, highlighting sensitivity to small-scale perturbations. It identifies the Lipschitz constant as a key robustness metric, deriving an analytical expression to estimate and enhance robustness based on network architecture.

Exploring the Potential of ChatGPT-4 for Clinical Decision Support in Cardiac Electrophysiology and Its Semi-Automatic Evaluation Metrics

This study enhances GPT-4’s domain expertise using the Retrieval-Augmented Generation (RAG) approach to integrate evidence-based knowledge. It also develops a semi-automatic evaluation framework combining metrics like BERT score, BLEURT, and cosine similarity with human assessments to streamline evaluation in medical decision-making.

Improve Certified Training with Signal-to-Noise Ratio Loss to Decrease Neuron Variance and Increase Neuron Stability

This work addresses over-regularization in certified training by introducing neuron variance and stability, analyzing their impact on robustness. Extending the Signal-to-Noise Ratio (SNR) framework, it proposes SNR-inspired loss functions to mitigate over-regularization and improve robustness.

Online Model Adaptation with Feedforward Compensation

This work introduces an online adaptation method with feedforward compensation, optimizing models using critical data from a memory buffer. It outperforms previous methods in slow time-varying systems, offering improved accuracy and robustness with a smaller error bound.

BioSLAM: A Bioinspired Lifelong Memory System for General Place Recognition

This paper presents BioSLAM, a lifelong SLAM framework with a dual-memory mechanism: dynamic memory for learning new observations and static memory for balancing old and new knowledge. It is the first memory-enhanced system for incremental place recognition in long-term navigation.

Proactive Human-Robot Co-Assembly: Leveraging Human Intention Prediction and Robust Safe Control

This paper introduces a framework for proactive human-robot collaboration, featuring intention prediction with human-in-the-loop training and robust safe control for interactive safety. The framework is demonstrated on a co-assembly task with a Kinova Gen3 robot.

Semi-Supervised Transfer Learning with Hierarchical Self-Regularization

This paper proposes a hierarchical self-regularization mechanism for transfer learning in semi-supervised settings, combining individual- and population-level regularizers with adaptive weighting based on sample confidence. The method also improves performance in fully-supervised and self-supervised tasks.

An Optical Control Environment for Benchmarking Reinforcement Learning Algorithms

This paper introduces an optical simulation environment for reinforcement learning-based controllers, modeling nonconvexity, nonlinearity, and time-dependent noise in optical systems. Benchmark results for various reinforcement learning algorithms are also provided.

General Place Recognition Survey: Towards the Real-World Autonomy Age

This paper surveys state-of-the-art methods in long-term localization, offering insights into key advancements and future opportunities. It serves as a tutorial for newcomers and a resource for advancing long-term robotics autonomy.

Property-Aware Relation Networks for Few-Shot Molecular Property Prediction

This paper proposes Property-Aware Relation (PAR) Networks for few-shot molecular property prediction, combining property-aware embeddings and adaptive relation graph learning to tailor molecular relations to specific properties. A meta-learning strategy selectively updates parameters to balance generic and property-specific knowledge.

Adaptive Consistency Regularization for Semi-Supervised Transfer Learning

This work introduces adaptive consistency regularization to unify semi-supervised and transfer learning, leveraging pre-trained source models and labeled/unlabeled target data. By adaptively selecting examples for regularization, the method improves target task performance and complements existing approaches for additional gains.

Escaping the Big Data Paradigm with Compact Transformers

This paper challenges the notion that transformers require large datasets, demonstrating that with appropriate size and tokenization, they can match or surpass CNNs on small datasets. The Compact Convolutional Transformer (CCT) achieves 98% accuracy on CIFAR-10 with only 3.7M parameters, outperforming other transformers in efficiency and rivaling ResNet50 at 15% of its size.

Robust Nonlinear Adaptation Algorithms for Multitask Prediction Networks

This paper presents an online adaptable multi-task model for predicting human intention and trajectory. The proposed approach integrates EKF with exponential moving average filtering and a dynamic multi-epoch update strategy to enhance accuracy using feedback from observed trajectories.

Robust Online Model Adaptation by Extended Kalman Filter with Exponential Moving Average and Dynamic Multi-Epoch Strategy

This paper addresses the train-test generalization gap by introducing online adaptation methods inspired by the Extended Kalman Filter (EKF) to update neural network models in test time.

Adaptable Human Intention and Trajectory Prediction for Human-Robot Collaboration

This paper proposes a multi-task model for predicting human trajectories and intentions, offering flexibility and adaptability for rapid integration into human-robot collaboration systems.

Automatic Generation of Personalized Comment Based on User Profile

This paper introduces the Personalized Comment Generation Network, a method for automatically generating social media comments that are relevant, grammatically accurate, and aligned with individual personality traits and styles.

Service

Reviewer

• Journals:

  IEEE T-RO, IEEE RA-L, IEEE T‐CSVT, TMLR.

• Conferences:

  NeurIPS, ICML, ICLR, ICRA, CVPR, ECCV, ICCV, ACCV, WACV, AAAI.

Program Committee Member

• ACL: Student Research Workshop

Teaching

Teaching assistant for 18-661: Introduction to Machine Learning for Engineers (CMU). Instructor: Prof. Yuejie Chi and Prof. Beidi Chen.

Teaching assistant for 18-786: Introduction to Deep Learning (CMU). Instructor: Prof. Aswin Sankaranarayanan.

Teaching assistant for Optoelectronics (PKU). Instructor: Prof. Zhiping Zhou.

Side Research

Apart from my research in AI, I have also gained substantial experience in optics and lasers, particularly during my master's studies. Below are some of the papers I have contributed to that focus on topics outside of AI.

Reinforcement Learning based Robust Control Algorithms for Coherent Pulse Stacking

This work introduces SAC-SPGDM, a hybrid algorithm combining stochastic parallel gradient descent with the soft actor-critic reinforcement learning method, enabling fast and robust optical control for coherent stacking of 128 pulses as demonstrated in simulations.

Sub-ps Resolution Clock-Offset Measurement over a 114 km Fiber Link using Linear Optical Sampling

This work demonstrates sub-ps resolution clock-offset measurement using linear optical sampling with dual optical frequency combs over a 114 km fiber link, achieving a time deviation of 110 fs at 1 s and 22 fs at 100 s averaging. This advance supports precise time synchronization in long fiber link applications.

Coherent Stacking of 128 Pulses from a GHz Repetition Rate Femtosecond Yb:Fiber Laser

This work demonstrates the coherent stacking of 128 ultrashort pulses from a 0.95 GHz femtosecond Yb:fiber laser using a delay line and polarization control, marking the first successful implementation of stacking 128 pulses.

Side Projects

These include courseworks and side projects.

WeHelp: A Shared Autonomy System for Wheelchair Users

This project introduces WeHelp, a shared autonomy system for wheelchair users with three modes: autonomous following via visual tracking, and remote control or teleoperation activated through speech recognition for seamless assistance.

Python Modeling for Ultrafast Optics and Supercontinuum Generation

This project presents a user-friendly platform with a GUI for simulating pulse evolution in optical fibers, including laser pulse representation, fiber modeling, and nonlinear processes like four-wave mixing and supercontinuum generation. It is designed for users of all experience levels.

Efficient Method for Categorize Animals in the Wild

This work proposes an efficient method for categorizing wild animals using advanced image augmentation (cutout, mixup, label smoothing) and ensemble learning. These techniques enhanced model robustness and secured a top 7/336 placement in the iWildCam 2019 competition at CVPR 2019.

Text Recognition in Noisy and Blurry Photographed Documents

This work presents an efficient pipeline for text recognition in noisy, blurry images by combining CTPN with DenseNet for detection and recognition. Fuzzy matching aligns detected text with a predefined library for refined results.

Ultimate Reversi (黑白棋)

This C++ Reversi (Othello) program includes a GUI and supports saving/loading games, online multiplayer, and AI battles. The AI is implemented with Alpha-Beta Pruning, Iterative Deepening, and a Translation Table for optimal performance.

Personal Stuff

I am deeply fascinated by Machine Learning, driven by the dream of achieving General AI to enhance life for everyone. Physics, especially Optics, captivates me with its profound mysteries.

Outside of work, I enjoy football (Soccer), traveling, and playing board games.

Along life’s road I have always sought truth,
In the search for verity, thought was always my guide.
My heart yearned without end for a chance of expression,
And longed to find words of meaning and grace.

Design and source code from Jon Barron's website