Research
I want robots to acquire real-world-facing geometric and physical common sense. My current work spans three mutually reinforcing threads — VLA models for contact-rich manipulation, RL for dynamic navigation, and end-to-end legged locomotion plus deployment.
1 · Geometric Failure Modes & Privileged-Geometry Distillation in VLA
Tsinghua SIGS · Oct 2025 – Present · In Preparation
I observe that current vision-language-action (VLA) models — pi0.5, TA-VLA, and friends — frequently fail on contact-rich and placement tasks because of occlusion and depth ambiguity: the trajectory looks roughly right, but the end-effector loses critical geometric cues in the last few centimetres.
What I'm working on:
- Collecting proprioception + dual D405 wrist + head D435 multi-modal data on a bimanual Songling + Jetson AGX Orin platform, covering optical-module insertion and block placement — tasks with sharp contact-sensitivity.
- Building action-conditioned 2.5D interaction maps, treating action-conditioned geometry as privileged information and distilling it into RGB-D-only deployment models.
- Systematically comparing diffusion-policy / flow-matching / autoregressive / hybrid VLA architectures on high-frequency action generation under contact.
Goal: a first-author submission to a top embodied-AI venue, evaluated by real-robot success rates on contact-rich tasks.
2 · RL Local Planning under Dynamic Obstacles
Eastern Institute of Technology, Ningbo · Jan 2026 – Present · Targeting T-RO / TMECH
Classical DRL-DCLP work targets static obstacles. The collaboration I joined pushes it into dynamic, mixed-density scenes:
- MDP redesign: re-defining state space and reward for varying-footprint robots under dynamic obstacles.
- 8-stage curriculum: progressively harder StageRos environments — adjacent stages scale map size by 0.8×, raise static density, and parameterise dynamic-obstacle counts.
- Temporal dynamic encoding: encoding the temporal state of multiple obstacles and fusing it into both policy and value networks for robustness in cluttered, time-varying environments.
3 · End-to-End Quadruped Locomotion & LiDAR Navigation
XJU Innovation Lab · Jun 2025 – Present · ICIC 2026 Oral, Accepted
A full sim-to-real loop:
- PPO and N-P3O policies in Isaac Gym / Isaac Lab for multi-terrain locomotion, generalising across ≥ 3 terrain types.
- On a self-designed quadruped, MuJoCo sim-to-sim then real-robot sim-to-real.
- Sustains > 10 N pushes with stable gait — walking 0.8 m/s, jogging 2 m/s, crawling — across 4 indoor / outdoor terrain classes and slopes.
- Indoor/outdoor SLAM with fast_lio / point_lio; ROS 2 Nav2 with A* / Dijkstra global planning and AMCL relocalisation.
Outcome: 1 first-author paper at ICIC 2026 (Oral, Accepted).
4 · Earlier Work · Deep-Learning Strawberry Pest Monitoring
Xinjiang University · Jun 2024 – Jan 2026 · Closed
YOLOv8 + attention + improved FPN: small-object recall +15 %, precision +25 %; pruned and quantised for embedded deployment. Coordinated rover + arm IK control. Closed as Outstanding under the National Innovation & Entrepreneurship Programme; one design patent (1st author) and one software copyright (3rd author).
What's Next
- First half of 2026: complete experiments on the VLA geometric-distillation method and submit my first-author paper; build a strong baseline for the RL dynamic-navigation work.
- Second half of 2026: graduate-school applications and continued research in embodied AI, with a focus on contact-aware perception × geometric reasoning.
If your work resonates, send me an email: 13325905201@163.com.