Litian Gong 「龚利天」

A sim-to-real pipeline for vision-language-action (VLA) models. We align a simulated DROID-style Franka setup with the real robot (controller, cameras, lab scene), render RL expert trajectories in simulation for distillation into π0.5-DROID, and LoRA-fine-tune π0.5 on real Franka FR3 data with out-of-distribution real-robot evaluation.

A controlled study of memory mechanisms for robot policies: published memory designs are re-implemented as plug-in modules on one shared Diffusion Policy backbone and evaluated under the same benchmark and protocol, including ablations where the policy's memory is masked, shuffled, or zeroed.

A modular world-model framework that breaks down design choices from driving, manipulation, and navigation world models into swappable components for controlled comparison. A learned latent world model is used as the internal simulator for model-predictive control in CarRacing (left: real environment, right: imagined future frames).

Learning-based person following for mobile robots, from a Unitree Go2-W wheeled quadruped to a Rosmaster X3 omnidirectional car. We build Isaac Sim / Isaac Lab environments with simulated humans and RTX lidar, replay real-robot runs as closed-loop digital twins running the real perception and decision stack, and train terrain-aware following policies on 3D street scenes.