Just Be Exemplary

Ph.D. Research Assistant in Robotics
Department of Computer Science
Clarkson University
Potsdam, New York, United States
Email: contact (at)
Photo of Xinchao Song
Home  Teaching  Publications  Projects  Certifications


Reinforcement-Learning Based Robotic Assembly of Fractured Objects Using Visual and Tactile Information

Xinchao Song*, Nikolas Lamb*, Sean Banerjee, Natasha Kholgade Banerjee. 2023 International Conference on Automation, Robotics and Applications (ICARA). [Paper]

We present a reinforcement learning approach that combines visual and tactile information to automatically assemble repair parts to fractured objects. We propose two novel visual metrics, which we term pixel offset error and assembly error, to provide estimation of assembly state. Our approach does not place constraints on object geometry and estimates the assembly state of the constituent objects in real time. We show tightly assembled fractured and restored pairs in simulation and on real robots.


Internet of Robotic Things: Current Technologies, Applications, Challenges and Future Directions

Davide Villa, Xinchao Song, Matthew Heim, Liangshe Li. arXiv preprint:2101.06256. [Paper]

The concept of the Internet of Things (IoT) is becoming increasingly popular, with the number of connected devices reaching billions. This paper focuses on the fusion of IoT and robotics, called the Internet of Robotic Things (IoRT), and discusses IoRT concepts, architectures, use case examples, key challenges, ethical issues, regulations, and future vision. This paper aims to provide a better understanding of the emerging concept of IoRT, its benefits and limitations, as well as guidelines and directions for future research and studies.


Belief-Grounded Networks for Accelerated Robot Learning under Partial Observability

Hai Nguyen*, Brett Daley*, Xinchao Song, Chistopher Amato, Robert Platt. 2020 Conference on Robot Learning (CoRL). [website]

We propose a method for policy learning under partial observability called the Belief-Grounded Network (BGN) in which an auxiliary belief-reconstruction loss incentivizes a neural network to concisely summarize its input history. Since the resulting policy is a function of the history rather than the belief, it can be executed easily at runtime. We compare BGN against several baselines on classic benchmark tasks as well as three novel robotic touch-sensing tasks. BGN outperforms all other tested methods and its learned policies work well when transferred onto a physical robot.


Imitation Learning in POMDPs with Contacts

Hai Nguyen, Xinchao Song, Christopher Amato, and Robert Platt. 2020 Robotics: Science and Systems (RSS): Reacting to Contact Workshop. [pdf]

We use the access to full states during training and imitation learning to train an intelligent agent that uses contacts to explore to solve a manipulation task. We formulate the problem as a partially observable decision process (POMDP) and solve it using imitation learning. We train a POMDP expert that solves the task while performing informative actions using contacts as well as an agent that acts on partial information by cloning this expert’s behavior. We test our method on a novel robotics domain and set up an experiment with a real robot.