research

Teaching robots to build

I work on embodied intelligence: getting robots to plan and build in the physical world. I started in autonomous-vehicle simulation, spent two years on robotic assembly at CMU's Intelligent Control Lab, and now work on reinforcement learning for vision-language-action models at UT Austin.

5 publications 27+ citations Google Scholar ↗

Illustration of physics-aware combinatorial assembly: a valid brick placement vs. placements that collapse or exhaust inventory — Valid vs. physically invalid placements under inventory constraints (Fig. 1 of the paper)

RA-L · IROS 2025 2025

Physics-Aware Combinatorial Assembly Sequence Planning using Data-free Action Masking

Ruixuan Liu, Alan Chen, Weiye Zhao, Changliu Liu

In what order should a robot place Lego-like primitives so the build never collapses halfway? We train a deep-RL construction policy guided by an online, data-free action mask that filters physically invalid moves before they happen. The planner sequences 250+ 3D structures with a 100% success rate; the best comparable baseline fails on more than 40 of them.

✦ Work done at the CMU Intelligent Control Lab, while I was in high school.

arXiv IEEE Xplore code

deep RLaction maskingassembly planning

AssemblyComplete: 3D Combinatorial Construction with Deep Reinforcement Learning: incomplete — The completion pipeline: library matching → scaled registration → constrained PPO construction

MECC 2025 2025

AssemblyComplete: 3D Combinatorial Construction with Deep Reinforcement Learning

Alan Chen, Changliu Liu

Hand a robot a half-built chair, and it should figure out what you meant to build, then finish it. AssemblyComplete matches the incomplete structure to an object library with point-cloud registration (RANSAC + ICP), then a PPO agent with an action-validity mask completes the assembly under stability, connectivity, and inventory constraints.

✦ First-authored in high school; funded in part by an IFAC student grant.

arXiv DOI code

deep RLpoint cloudsassembly completion

System architecture connecting SUMO simulation, FLOW, and RLlib policy training — The SUMO ↔ FLOW ↔ RLlib training loop behind the eco-driving policies

Science Progress 2024 2024

Developing an Eco-Driving Strategy in a Hybrid Traffic Network Using Reinforcement Learning

Umar Jamil, Mohammadreza Malmir, Alan Chen, Monika Filipovska, Mimi Xie, Caiwen Ding, Yu-Fang Jin

Can a handful of learning-controlled autonomous vehicles calm an entire city grid? We model a Manhattan-style network in SUMO + FLOW and train RLlib agents whose eco-driving policies smooth traffic flow and cut fuel consumption in mixed human/autonomous traffic.

DOI (open access)

autonomous vehicleseco-drivingtraffic networks

The learned policy assembling a structure on the real FANUC arm (8x)

arXiv 2023 2023

Simulation-Aided Learning from Demonstration for Robotic LEGO Construction

Ruixuan Liu, Alan Chen, Xusheng Luo, Changliu Liu

A human demonstrates a LEGO build; the robot learns the manipulation policy safely in simulation, then reproduces the build in the real world. I worked on the simulation-aided learning-from-demonstration pipeline and its Sim2Real deployment on a FANUC LR Mate arm.

arXiv

learning from demonstrationsim2realdigital twin

ICIBM 2022 & Tran-SET 2022 2022

Interactive Data Collection Using CARLA and OpenCDA for Reinforcement Learning

Alan Chen, Joseph Clemmons, Umar Jamil, Ashley Land, Sara Ahmed, Yu-Fang Jin

My first research project, the summer after freshman year of high school: interactive data-collection tooling on CARLA + OpenCDA co-simulation, so RL agents for connected autonomous vehicles could train on realistic interaction data.

program

CARLAdata collectionconnected AVs

in progress

Current work at LARG: fine-tuning robot foundation models (Octo, OpenVLA) with PPO/GRPO to learn mobile manipulation from scratch, and independent exploration of kernel optimization and RLMs. If you want to chat, email me.

example of a policy rollout · ManiSkill