HETIAN GUO ROBOT LEARNING
ROBOT LEARNING / RESEARCH PORTFOLIO
I am Hetian Guo, a Ph.D. student in Artificial Intelligence at Jilin University. My research focuses on VLA policy learning, predictive world models, visuo-tactile manipulation, and whole-body motion generation, with experiments on bipedal humanoids, wheeled platforms, and bimanual setups.
From predicted latent states to cross-platform robot experiments
01 / SELECTED WORK
Predictive future states, temporal belief, tactile contact dynamics, and whole-body motion generation.
PLaW-VLA: Predictive Latent World Modeling for Vision-Language-Action Policies
Co-first author
Predicts future states in a pretrained visual latent space rather than reconstructing pixels and couples state prediction with action generation. Average success reaches 97.4% on LIBERO and 72.7% on LIBERO-Plus, with long-horizon, fine-grained, and deformable-object manipulation tested on Astribot S1.
BeliefVLA: Implicit Belief-State Driven Vision-Language-Action Models for Robust Manipulation
Second author
Infers an implicit belief state from observation history to add temporal memory and dynamics priors to VLA control; evaluated bimanual tasks and visual perturbations on Realman, ARX R5, and AgileX PIPER.
SeMoCo: A Semantic-First Motion Codec for Motion Language Modeling
Co-first author
A semantic VQ and kinematic-residual RVQ form hierarchical motion tokens; a dual-axis generator models temporal semantics and refines motion within each token. The work also builds the unified-skeleton Ω-MotionVerse dataset, with generated motions replayed on G1 through Sonic.
TacDyn-WAM: Learning Implicit Tactile Dynamics in a Heterogeneous Visuo-Tactile World Action Model
Collaborative research
Learns a dynamics-aware tactile target space to predict contact evolution at multiple horizons and couples visual/tactile experts for action generation; achieved 81.5% average success on UniVTAC with real-robot validation.
02 / REAL ROBOTS
Bipedal humanoids, wheeled platforms, and cross-platform bimanual manipulation.
Sonic motion execution, Inspire dexterous-hand integration, multi-task robot data collection, and SeMoCo motion replay.
BeliefVLA perturbation tests on Realman; PLaW-VLA long-horizon and fine-grained manipulation on Astribot S1.
Cross-platform bimanual tasks for BeliefVLA, probing how observation history and implicit belief states support robust control.
03 / BACKGROUND
Ph.D. research in artificial intelligence, with work spanning robot learning and spatiotemporal modeling.
Deployed Sonic and GMR + AMO motion pipelines on Unitree G1 and integrated Inspire hands; collected multi-task teleoperation data and fine-tuned/deployed Psi-0 and GR00T N1.7. Studied SeMoCo whole-body motion generation (AAAI 2027 submission) and replayed generated motion on G1 through Sonic.
Studied waterway-network extraction, trend/seasonality decomposition, and static/dynamic spatial relations in North Sea and Baltic Sea AIS trajectories; contributed to a marine traffic forecasting model published in Transportation Research Part E (2025).
Ph.D. in Artificial Intelligence
Research: embodied AI, VLA, world-action models, dexterous-hand control; Advisor: Prof. Zipei Fan
M.S. in Engineering (Mechanical Engineering)
Research: intelligent transportation, autonomous driving, hardware-software co-simulation; Advisor: Prof. Yunli Shao
B.Eng. in Computer Science and Technology
Research: intelligent transportation, spatiotemporal data mining, traffic forecasting; Advisor: Prof. Xuan Song
04 / INTELLIGENT TRANSPORTATION SYSTEMS
Prior work on traffic-agent trajectory prediction and spatiotemporal marine traffic modeling.
First author · Heterogeneous higher-order interactions for grouped traffic-agent trajectories
Second author · AIS waterway networks and traffic forecasting; research from the LocationMind internship
05 / RESEARCH & ENGINEERING
Research methods, robot-system practice, and software implementation grounded in the work above.
VLA / world-action models, predictive latent states, visuo-tactile representations, language-conditioned whole-body motion, and policy fine-tuning.
Experiments on bipedal humanoids, wheeled platforms, and bimanual setups; multi-task teleoperation data collection, motion-pipeline deployment, and replay.
Build training and evaluation pipelines with Python / PyTorch, integrate robot control with C++ and ROS / ROS2, and use Transformers and Diffusers in generative-model experiments.
06 / CONTACT
I welcome conversations about robot learning, VLA, world models, and research internships.