Robo360: A 3D Omnispective Multi-Material Robotic Manipulation Dataset

Litian Liang, Liuyu Bian, Caiwei Xiao, Jialin Zhang, Linghao Chen, Isabella Liu, Fanbo Xiang, Zhiao Huang, Hao Su

Preprint

Keywords: 1. Real-World Dataset, 2. Robotic Manipulation

Paper / Video / Teleoperation / Code and data coming soon

We collected a large-scale real-world dataset of robot manipulation of objects with diverse optical and material variations from omnispective view directions using 86 DSLR cameras. We demonstrated the quality of Robo360 dataset by showing you can not only learn robot policies with imitation learning but also achieve high-quality reconstruction of the scene in 3D. We hope that Robo360 can open new research directions yet to be explored at the intersection of understanding the physical world in 3D and robot control.

Work done at Hao Su Lab @ UC San Diego

Reparameterized Policy Learning for Multimodal Trajectory Optimization

Zhiao Huang, Litian Liang, Zhan Ling, Xuanlin Li, Chuang Gan, Hao Su

ICML 2023 Oral Presentation

Keywords: 1. continuous action space, 2. dense reward with local optima, 3. sparse reward exploration

Paper / Code / Project Page / Video

We proposed a novel policy parameterization that significantly improves performance in maximizing with a dense reward that contains local optima and exploring sparse reward (sparse 0/1 reward) with intrinsic rewards (RND). Our novel approach solves hard sparse reward exploration tasks that were previously believed to be impossible to solve by RL methods.

Work done at Hao Su Lab @ UC San Diego

Reducing Variance in TD Value Estimation via Ensemble of Deep Networks

Litian Liang, Yaosheng Xu, Stephen McAleer, Dailin Hu, Alexander Ihler, Pieter Abbeel, Roy Fox

ICML 2022

Keywords: 1. TD value estimation bias and variance, 2. ensemble value estimator

Paper / Code

We investigate and propose a specific combination of existing techniques to improve sample efficiency in learning a value network with TD. The combination achieved enough stability during training to obviate the target network. This method is applicable to any value network that optimizes TD error.

Work done at Intelligent Dynamics Lab @ UC Irvine

University of California, San Diego

M.S. in Computer Science (Thesis Track) [Sep. 2022 - Dec. 2023]

Research Advisor: Prof. Hao Su, PhD candidate Zhiao Huang

Research Topics: 

1. 3D vision for Embodied AI

2. Latent variable models, Model-Based RL, Sparse Reward Exploration

3. Differentiable Simulation of Dynamical Systems

University of California, Irvine

B.S. in Computer Science (Honors Program) [Sep. 2018 - Mar. 2022]

Research Advisor: Prof. Roy Fox, Prof. Alexander Ihler

Research Topics: 

1. Temporal-Difference value estimation, Maximum Entropy RL

2. Domain Randomization for Sim2Real Transfer