Dian Chen 陈典

Hi! I am a machine learning researcher at Apple AI/ML on the AFM team, working on foundation model. Previously I worked on self-driving at Waabi. I obtained my Ph.D in CS at UT Austin, under the supervision of Prof. Philipp Krähenbühl . During this time, I interned at Waymo, working on multi-agent behavior prediction.

Email  /  GitHub  /  Scholar

We propose DriveGATr, a novel architecture for modeling agents that achieves SE(2)-equivariance without the computational cost of existing methods, leveraging geometric deep learning

We present MotionLM, a behavior predictor that represent continuous trajectories as sequences of discrete motion tokens. MotionLM casts multi-agent motion prediction as a language modeling task over this domain.

We introduce COOPERNAUT, an end-to-end learning model that uses cross-vehicle perception for vision-based cooperative driving. Our model encodes LiDAR information into compact point-based representations that can be transmitted as messages between vehicles via realistic wireless channels.

We present LAV, a mapless, learning-based end-to-end driving system. LAV takes as input multi-modal sensor readings and learns from all nearby vehicles in the scene for both perception and planning. At test time, LAV predicts multi-modal future trajectories for all detected vehicles, including the ego-vehicle. Our system outperforms all prior methods on the public CARLA Leaderboard by a wide margin, improving driving score by 25 and route completion rate by 24 points.

We present a model-based RL method for autonomous driving and navigation tasks. The world model is factorized into a passively moving environment, and a compact ego component. Our method significantly simplifies reinforcement learning. It ranks first on the CARLA leaderboard, and outperforms state-of-the-art imitation learning and model-free reinforcement learning on driving tasks. It is also an order of magnitude more sample efficient than model-free RL on the navigation games in the ProcGen benchmark.

We present a two-stage imitation learning method for vision-based driving. Our approach achieves 100% success rate on all tasks in the original CARLA benchmark, sets a new record on the NoCrash benchmark, and reduces the frequency of infractions by an order of magnitude compared to the prior state of the art.

We present a robotic system that learns to segment its visual observations into individual objects by experimenting with its environment in a completely self-supervised manner. Our system is at par with the state-of-art instance segmentation algorithm trained with strong supervision.

We present a novel skill policy architecture and dynamics consistency loss which extend visual imitation to more complex environments while improving robustness. Experiments results are shown in a robot knot tying task and a first-person visual navigation task.

We present a system where a robot takes as input a sequence of images of a human manipulating a rope from an initial to goal configuration, and outputs a sequence of actions that can reproduce the human demonstration, using only monocular images as input.

template