Jun 1 – Aug 31, 2026 · Honor Device Co., Ltd. · Summer internship
Honor Humanoid: Online RL for VLA
VLA Algorithm Engineer Intern at Honor. Built a human-in-the-loop online SFT / online RL pipeline for Honor's self-developed humanoid robot — smooth teleoperator takeover, an intervention-driven data pipeline, and a distributed Actor–Learner–Robot learning loop.
Working on Honor’s self-developed humanoid robot: teleoperation data collection, replay, training, and on-robot deployment, along with online RL algorithm research for Vision-Language-Action (VLA) models on custom manipulation tasks.
Demos
On-robot inference under perturbations
Autonomous pick-and-place rollouts on the real robot while the object (left) and the target basket (right) are perturbed mid-episode — the policy re-tracks and completes the task.
Data design for robustness
Four categories of teleoperated demonstrations: standard demonstrations, perturbations during the pick phase, perturbations during the place phase, and recovery from failed-grasp states — designed so the policy learns to recover, not just to repeat.
Online SFT: before vs. after
A task the offline-trained policy could not perform is learned on the robot through the online learning loop — human takeover data flows back into training, and the takeover ratio drops as new checkpoints are dispatched.
What I built
Smooth human takeover. Seamless human-teleoperator intervention during autonomous rollouts, including the autonomy–teleoperation transition state machine, so a teleoperator can take over and hand back control mid-episode without disturbing the robot.
Intervention-driven data pipeline. On-robot data recording during takeovers, full local disk persistence, and automatic upload to the development server for training.
Distributed learning loop. A closed loop across three tiers: the robot’s onboard AGX/NX boards (topic publishing, command execution, data collection), an x86 host as the Actor (model inference, streaming actions to the robot), and a development server as the Learner (data aggregation, training, checkpoint dispatch) — a complete online SFT / online RL pipeline.
Algorithm research
Researching RECAP-style (π0.6) offline RL combined with noise-space fast adaptation (DSRL-like): learning corrective noise from human-intervention data so the policy can rapidly imitate takeover behaviors before a new checkpoint finishes training — substantially raising success rates on tasks uncovered by offline training — and further adapting the algorithm to loco-manipulation tasks.