Jingxi Lu

Research

Learning systems grounded in dynamics and safety.

I study how robots can continue to move and act safely when actuators, terrain, and observations depart from nominal conditions. My work spans reinforcement learning, safety-critical control, real-world robot integration, and efficient multimodal inference.

Featured · Ongoing at HIER Lab

Fault-Tolerant Quadruped Locomotion

Actuator weakness, lock-up, and loss of torque change the forces and motions a robot can physically realize. I develop fault-conditioned locomotion policies and safety filters that account for these changes instead of relying on explicit fault detection alone.

The current framework combines curriculum-based reinforcement learning with actuator torque-speed envelopes, floating-base dynamics, and control barrier functions. A quadratic program minimally modifies commands while enforcing feasible contact forces, joint limits, and roll and pitch safety conditions on the Unitree Go2.

  • PPO
  • CBF / HOCBF
  • Quadratic programming
  • Isaac Lab
  • MuJoCo
HIER Lab

IEEE RA-L · ICRA 2026

STATE-NAV

STATE-NAV reframes traversability for bipedal robots as a stability-aware command velocity rather than a geometric score. TravFormer predicts locomotion instability from terrain and commanded motion, then supplies a hierarchical planner with the fastest command that remains below a user-defined risk limit.

I contributed to the real-world Digit deployment, including ROS1/ROS2 LiDAR and pose integration, robot-centric elevation maps, and low-latency inference for the traversability network.

  • Digit
  • ROS1 / ROS2
  • LiDAR
  • Elevation mapping
  • Transformer inference
Project page Paper Video

RedVTP · HIVTP

Efficient Vision-Language Model Inference

I worked on training-free visual-token pruning for autoregressive and diffusion vision-language models. HIVTP uses middle-layer attention and hierarchical global-local selection to retain fine-grained information. RedVTP uses attention from still-masked response tokens to prune visual tokens after the first diffusion inference step.

Across the evaluated diffusion VLMs, RedVTP reduced inference latency by up to 64.97% and improved token-generation throughput by up to 186% while preserving—and sometimes improving—benchmark accuracy.

  • PyTorch
  • Hugging Face
  • Diffusion VLMs
  • Attention analysis
  • lmms-eval
RedVTP paper RedVTP code HIVTP paper HIVTP code
64.97% lower latency on LLaDA-V
186% higher token-generation throughput

Additional workEarlier research threads.

Projects that shaped my experience in robot learning, systems optimization, and sequential decision-making.

Robotic learning

Humanoid and Bipedal Locomotion

PPO locomotion for the Hector biped and motion-retargeting and behavior-cloning pipelines for Unitree G1 whole-body imitation.

IEEE TMC 2026

Hybrid Microservice Scheduling

Behavior cloning and Soft Actor-Critic for cold-start-aware scheduling under dynamic edge-resource constraints.

Paper Code

ICCSIE 2025

RL-Initialized Column Generation

PPO, graph attention, and pointer networks for generating strong initial columns in aircraft-recovery optimization.

DOI