Graduate Summary @CMU
Good morning! I am a fourth year PhD student at CMU's Robotics Institute. I am working with Professor Maxim Likhachev on search-based planning. I am specifically interested in combining search with machine learning and in multi-agent planning. I am always happy to discuss research and brainstorm ideas, and am interested in research collaborations -- feel free to email me!
Undergraduate Summary @UC Berkeley
I graduated U.C. Berkeley in spring 2020 with a B.S. in Electrical Engineering & Computer Science and a B.A. in Applied Mathematics.
Apart from classes, I spent a significant part of my undergraduate career doing research and teaching. I conducted research with Professor Sergey Levine in Berkeley AI Research, and was a teaching assistant in several classes (EE16A, CS188, CS170 x2).
In the past, I have worked at Two Sigma, Lawrence Livermore National Laboratory with Professors Gerald Friedland and Kannan Ramchandran, and the Laboratory of Quantitative Imaging at Stanford University with Professor Daniel Rubin.
Teaching
I was very active in teaching at UC Berkeley and thoroughly enjoyed the experience. I am particularly interested in improving myself as a teacher and improving the internal (teaching assistant facing) structures of courses. Please see my teaching page for more information!
Effective Integration of Weighted Cost-to-go and Conflict Heuristic within Suboptimal CBS
Rishi Veerapaneni, Tushar Kusnur, Maxim Likhachev
arXiv Full version / SoCS 2022 Extended Abstraction version / AAAI 2023 Version
Conflict-Based Search (CBS) is a popular multi-agent path finding (MAPF) solver that employs a low-level single agent planner and a high-level constraint tree to resolve conflicts. The majority of modern MAPF solvers focus on improving CBS by reducing the size of this tree through various strategies with few methods modifying the low level planner. All low level planners in existing CBS methods use an unweighted cost-to-go heuristic, with suboptimal CBS methods also using a conflict heuristic to help the high level search. Contrary to prevailing beliefs, we show that the cost-to-go heuristic can be used significantly more effectively by weighting it in a specific manner alongside the conflict heuristic. We introduce two variants of doing so and demonstrate that this change can lead to 2-100x speedups in certain scenarios. Additionally, we show the first theoretical relation of prioritized planning and bounded suboptimal CBS and demonstrate that our methods are their natural generalization.
Minimizing Coordination in Multi-Agent Path Finding with Dynamic Execution
Aidan Wagner*, Rishi Veerapaneni*, Maxim Likhachev
Eighteenth AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment (AIIDE), 2022
Modern Multi-agent Path Finding (MAPF) solvers plan assuming that agents directly follow the space-time trajectories at known constant speeds without delays or speedups, resulting in rigid plans which need to be replanned if there are changes during execution. Instead we would like agents to be able to follow their computed paths with dynamic velocities while requiring minimal coordination with others to prevent collisions and deadlocks. We introduce a novel paradigm and show how planning in space-coordination level, rather than space-time, allows us to simultaneously plan paths and a coordination controller. Our method, Space-Level Conflict-Based Search (SL-CBS), builds on the Conflict-Based Search framework and allows us to reason explicitly about coordination, producing paths as well as a coordination controller with bounded suboptimal minimal coordination. We show experimentally that this results in a 20-50% reduction in coordination compared to the closest state of the art solver.
Entity Abstraction in Visual Model-Based Reinforcement Learning
Rishi Veerapaneni*, John D. Co-Reyes*, Michael Chang*, Michael Janner, Chelsea Finn, Jiajun Wu, Joshua B. Tenenbaum, Sergey Levine
Entity Abstraction in Visual Model-Based Reinforcement Learning. Conference on Robot Learning (CoRL), 2019
Paper / Project webpage / Code
We introduced a framework for model-based planning that predicts and plans with learned object representations without supervision. The key idea behind our approach is to frame model-based planning under the language of a factorized HMM that processes a set of hidden states independently and symmetrically. This approach gives us permutation invariance, order invariance, and count equivariance by collapsing the combinatorial complexity along the object dimension. We show on a combinatorially complex block-stacking task that we are able to achieve almost three times the accuracy of non-latent-factorized video prediction model and outperform an oracle model that assumes access to object segmentations.
This work also appeared in:
Object Abstraction in Visual Model-Based Reinforcement Learning. Perception as Generative Reasoning (PGR) workshop, NeurIPS 2019
Tricking Neural Networks: Create your own Adversarial Examples
Daniel Geng, Rishi Veerapaneni
Article published on January 10, 2018 at ML@Berkeley
A fifteen minute read that introduces the concept of adversarial examples and how to construct them (figuratively and literally). We walk the readers through code snippets that show how different types of adversarial examples can be created.
Adaptive Estimation of Active Contour Parameters Using Convolutional Neural Networks and Texture Analysis
Assaf Hoogi, Arjun Subramaniam*, Rishi Veerapaneni*, Daniel Rubin
Adaptive estimation of active contour parameters using convolutional neural networks and texture analysis, IEEE Transactions on Medical Imaging, vol. 36, no. 3, March 2017
We created a generalization of the level set segmentation approach by supplying a novel method for adaptive estimation of active contour parameters. The presented segmentation method is fully automatic once the lesion has been detected.