publications

Submitted and in preparation

2021

1. Learning to Map for Active Semantic Goal Navigation Georgakis Georgios, Bernadette Bucher, Siddharth Singh , Karl Schmeckpeper, Kostas Daniilidis Submitted to the Robotics: Science and Systems Conference (RSS) 2021 [Abstract]

   — We consider the problem of object goal navigation inunseen environments. In our view, solving this problem requireslearning of contextual semantic priors, a challenging endeavourgiven the spatial and semantic variability of indoor environments.Current methods learn to implicitly encode these priors throughgoal-oriented navigation policy functions operating on spatialrepresentations that are limited in the agent’s observable areas.In this work, we propose a novel framework that actively learnsto generate semantic maps outside the field of view of the agentand leverages the uncertainty over the semantic classes in theunobserved areas to decide on long term goals. We demonstratethat through this spatial prediction strategy, we are able to learnsemantic priors in scenes that can be leveraged in unknownenvironments. Additionally, we show how different objectives canbe defined by balancing exploration with exploitation duringsearching for semantic targets. Our method is validated inthe visually realistic environments offered by the Matterport3Ddataset and show state of the art results on the object goalnavigation task.

Accepted papers

2020

1. CURIOSITY INCREASES EQUALITY IN COMPETITIVE RESOURCE ALLOCATION Bucher, Bernadette*, Singh, Siddharth*, Clelia de Mutalier, Kostas Daniilidis, Vijay Balasubramanian In Bridging AI and Cognitive Science (BAICS) workshop of The International Conference on Learning Representations 2020 [Abstract] [Link]

   We consider multiple agents using different strategies to compete for resources with a diverse distribution of rewards. Statistical theory shows that two kinds of equilibria are possible: (1) where some agents “settle” on a fixed resource while others visit diverse sites, and (2) where all agents pursue a similar strategy of visiting diverse sites. The first equilibrium shows a highly skewed reward distribution; in the second equilibrium most agents are similarly successful. We show that a population of agents can learn these equilibrium strategies through reinforcement learning. In conventional Q-learning, the population of agents learns the equilibrium strategy with skewed rewards. If we add curiosity, an intrinsic motivation to explore, Q-learning converges faster, and produces the second equilibrium in which most agents get similar average rewards. Thus, curiosity increases equality.

2019

1. RoboNet: Large-Scale Multi-Robot Learning Dasari,Sudeep , Ebert, Frederik , Tian, Stephen , Nair, Suraj , Bucher, Bernadette , Schmeckpeper, Karl , Singh, Siddharth, Levine, Sergey , Finn, Chelsea In Conference on Robot Learning (CoRL) 2019 [Abstract] [Link] [Website]

   Robot learning has emerged as a promising tool for taming the complexity and diversity of the real world. Methods based on high-capacity models, such as deep networks, hold the promise of providing effective generalization to a wide range of open-world environments. However, these same methods typically require large amounts of diverse training data to generalize effectively. In contrast, most robotic learning experiments are small-scale, single-domain, and single-robot. This leads to a frequent tension in robotic learning: how can we learn generalizable robotic controllers without having to collect impractically large amounts of data for each separate experiment? In this paper, we propose RoboNet, an open database for sharing robotic experience, which provides an initial pool of 15 million video frames, from 7 different robot platforms, and study how it can be used to learn generalizable models for vision-based robotic manipulation. We combine the dataset with two different learning algorithms: visual foresight, which uses forward video prediction models, and supervised inverse models. Our experiments test the learned algorithms' ability to work across new objects, new tasks, new scenes, new camera viewpoints, new grippers, or even entirely new robots. In our final experiment, we find that by pre-training on RoboNet and fine-tuning on data from a held-out Franka or Kuka robot, we can exceed the performance of a robot-specific training approach that uses 4x-20x more data. For videos and data, see the project webpage: https://www.robonet.wiki/




2. F1/10: An open-source 1/10th scale platform for autonomous racing and reinforcement learning O,Kelly,Matthew , Dhruv Karthik , Hongrui Zheng , Joseph Auckley , Siddharth Singh , Shashank D Prasad , Kim Luong, Matt R Lebermann , Rahul Mangharam In Neural Information Processing Systems (Neurips) 2019 [Abstract] [Link] [Website]

   The deployment of learning algorithms on autonomous vehicles is expensive, slow, and potentially unsafe. We present the F1/10 platform, a low-cost open-source 1/10th scale racecar and validated simulator which enables safe and rapid experimentation suitable for laboratory research settings. The F1/10 environment and hardware offer an accessible way to validate reinforcement learning policies learned in simulation and on a real car via the same intuitive OpenAI Gym API. To this effect, we demonstrate two approaches based on reinforcement learning that enable the transfer of policies from simulation to a real life track. The first demonstration details the use of our modular photorealistic simulator to learn an end-to-end racing policy in a self-supervised manner. Here, we demonstrate the F1/10’s capability for distributed online learning by sending batches of ‘experiences’ (video streams, odometry, etc.) back to a server, which asynchronously trains on this data and updates the racecar’s network weights on the fly. We also show a way to take quasi-deterministic ‘steps’ and perform ‘resets’ on the real car, thereby more closely mimicking the Gym API standards. The second demonstration uses our lightweight physics simulator to perform a joint optimization over a parameterized description of the racing trajectory, planning algorithms, and car dynamics, resulting in performance which exceeds all other entries in real-life races.

2018

1. Gradient Aware - Shrinking Domain based Control Design for Reactive Planning Frameworks used in Autonomous Vehicles Adarsh Modh* , Siddharth Singh* , A. V. S. Sai Bhargav Kumar , Sriram N. N. , K. Madhava Krishna In Advances in Robotics (AIR) 2018 [Abstract] [Link]

   — In this paper, we present a novel control law for longitudinal speed control of autonomous vehicles. The key contributions of the proposed work include the design of a control law that reactively integrates the longitudinal surface gradient of road into its operation. In contrast to the existing works, we found that integrating the path gradient into the control framework improves the speed tracking efficacy. Since the control law is implemented over a shrinking domain scheme, it minimizes the integrated error by recomputing the control inputs at every discretized step and consequently provides less reaction time. This makes our control law suitable for motion planning frameworks that are operating at high frequencies. Furthermore, our work is implemented using a generalized vehicle model and can be easily extended to other classes of vehicles. The performance of gradient aware - shrinking domain based controller is implemented and tested on a stock electric vehicle on which a number of sensors are mounted. Results from the tests show the robustness of our control law for speed tracking on a terrain with varying gradient while also considering stringent time constraints imposed by the planning framework.