Tao Chen
taoc1@andrew.cmu.edu

I am a master student (MSR) at Robotics Institute, Carnegie Mellon University, from fall 2017, advised by Abhinav Gupta. My research interests revolve around the intersection of learning, control, and manipulation.

Prior to this, I was one of the core AI Researchers and Engineers in Shanghai LingXian Robotics, where I conducted researches on object detection, image segmentation, deep reinforcement learning in robotics and control, SLAM, etc.

I earned my bachelor's degree from Shanghai Jiao Tong University in June, 2016, majoring in mechanical engineering and automation. I was also an exchange student (GEARE program) at School of Mechanical Engineering, Purdue University.

Email  /  CV  /  Google Scholar  /  GitHub

Research

Learning Exploration Policies for Navigation
Tao Chen, Saurabh Gupta, Abhinav Gupta
International Conference on Learning Representations (ICLR), 2019, under review
project page

Numerous past works have tackled the problem of task-driven navigation. But, how to effectively explore a new environment to enable a variety of down-stream tasks has received much less attention. In this work, we study how agents can autonomously explore realistic and complex 3D environments without the context of task-rewards. We propose a learning-based approach and investigate different policy architectures, reward functions, and training paradigms. We find that use of policies with spatial memory that are bootstrapped with imitation learning and finally finetuned with coverage rewards derived purely from on-board sensors can be effective at exploring novel environments. We show that our learned exploration policies can explore better than classical approaches based on geometry alone and generic learning-based exploration techniques. Finally, we also show how such task-agnostic exploration can be used for down-stream tasks.


Hardware Conditioned Policies for Multi-Robot Transfer Learning
Tao Chen, Adithya Murali, Abhinav Gupta
Advances in Neural Information Processing Systems (NeurIPS), 2018
arXiv / project page

Deep reinforcement learning could be used to learn dexterous robotic policies but it is extremely challenging to transfer them to new robots with vastly different hardware properties. It is also prohibitively expensive to learn a new policy from scratch for each robot hardware due to the high sample complexity of modern state-of-the-art algorithms. We propose a novel approach called Hardware Conditioned Policies where we train a universal policy conditioned on a vector representation of robot hardware. We considered robots in simulation with varied dynamics, kinematic structure, kinematic lengths and degrees-of-freedom. First, we use the kinematic structure directly as the hardware encoding and show great zero-shot transfer to completely novel robots not seen during training. For robots with lower zero-shot success rate, we also demonstrate that fine-tuning the policy network is significantly more sample efficient than training a model from scratch. In tasks where knowing the agent dynamics is crucial for success, we learn an embedding for robot hardware and show policies conditioned on the encoding of hardware tend to generalize and transfer well.


Development of a Soft Elastomeric Gripper for Dexterous Grasping
Tao Chen, Guo-Ying Gu
Bachelor's Thesis, 2016
Awarded 2016 Excellent Bachelor Thesis (Top 1%) of Shanghai Jiao Tong University
project page

Development of robotic grippers that can grasp universal objects analogous to humans hands is still a challenging and open problem in the robotics society. I designed a four-fingered soft gripper, in which each finger is a bending multi-cavity pneumatic elastomer actuator (MCPEA). The fingers are modularized and entirely composed of soft elastomer material that is inherent compliant. I also developed a finite-element analysis (FEA) model to analyze the kinematics of the MC-PEA. The influences of the geometric parameters on the bending performance in terms of the bending angle and output force at the distal tip are investigated as well. The four-fingered soft gripper is able to grasp universal unknown objects with different sizes and shapes even frangible objects, such as vegetables, fruits, eggs, noodles, pens, candles, electronics, and cups.

Work Experience

Automatic Exploration for Dense Map Building
Tao Chen
Shanghai LingXian Robotics, 2017
project page

A key step for robots to get popularized into our daily life is that robots should be able to automatically explore the new environment when they are deployed in new houses or buildings. In this project, I combine the strength of motion planning (OMPL and SBPL), frontier-based exploration, SLAM (ORB-SLAM2), and object recognition and segmentation (FCIS) techniques to build an automatic mapping system that can autonomously explore the new houses, recognize daily objects and remember their locations while keep building the dense map as it moves. After the map is built, the robot can be asked to find and move to the objects it has seen (like cup, monitor) autonomously.

Learning to Play Table Tennis with Deep Learning and Guided Policy Search
Tao Chen
Shanghai LingXian Robotics, 2016
video1 / video2 / code

This project was aimed at making a UR 10 robot able with stereo cameras to play table tennis with only visual inputs. It was mainly divided into two stages:

  • Use Highway Network + LSTM + MDN to predict the ball’s position (trajectory) in the future frames
  • Use GPS(Guided Policy Search) to train a UR 10 robot to reach the target poses which cover a high-dimensional configuration space for playing table tennis

Automatic Parallel Parking with Reinforcement Learning
Tao Chen
Shanghai LingXian Robotics, 2016
video1 / video2 / code

This project was aimed at making car (agent) learn parallel parking with reinforcement learning in Gazebo. The agent follows the Ackermann steering model and utilizes PID control for velocity control. A faster simulator was developed outside of Gazebo to speed up the training and the model was later transfered into the Gazebo envrionment. Q learning and imtation learning are the two principle techniques used here.

Course Projects

Design and Manufacturing of a Tennis Ball Collecting Robot
Tao Chen, Matthew Stouder, Zhedong Han, Duankang Fu, Sara Lyons, Zhishang Xu
GEARE program, Purdue University, 2015
video / code

This project was the Senior Engineering Design Capstone project at Purdue University. We built an aesthetically pleasing tennis ball collecting robot which can collect tennis balls dispersed on a tennis court. I was fully responsible for all the programming and control tasks for the robot.

2015 RoboMaster Robotics Competition
Tao Chen, Mechanical Team Leader
hosted by DJI , 2015
video

This competition was a real-life version of Counter-Strike game with real mobile robots. I led and managed the mechanical group. We won the second prize in 2015 RoboMasters national robot competition of east China and the third prize in 2015 RoboMasters national robot competition final.

An Active Rehabilitation Device for Elbow Joints
Tao Chen, Leader
Advisor: Prof. Hua Shao
Engineering Design, 2015
China Patent, CN105148460B
project page

We designed and built an inexpensive yet effective elbow joint rehabilitation device. The device is only composed of mechanical parts such as a lead screw, and a four-bar mechanism. It can help patients exercise their elbow joints in an inexpensive way, and it is also very easy to use and portable. We have applied a patent (Application Number: CN201510472161.5, Publication Number: CN105148460B) for this rehabilitation device.

A High-adaptability Rescue Robot
Tao Chen, Leader
Advisor: Prof. Qinghua Liang
Design and Manufacturing, 2014
code 1 / code 2

We built a high-adaptability track robot with two separate frames. I led and managed the team and I was responsible for electronic control, programming and part of manufacturing.


This guy makes a nice webpage.