I’m Bikram,

Bikram Pandit, Ashutosh Gupta, Mohitvishnu S. Gadde, Addison Johnson, Aayam Kumar Shrestha, Helei Duan, Jeremy Dao, Alan Fern
Conference on Robot Learning (CoRL), 2024

Payload transport via multi-wheel robot carriers is well-understood, effective, and configurable on flat terrain. This paper aims to achieve similar effectiveness and configurability for rough terrain using multi-biped robot carriers. We replace wheels with multiple bipedal robots attached to the carrier. Our main contribution is designing a decentralized controller for such systems that works with varying numbers and configurations of these robots without retraining. We employ reinforcement learning to train the controller in simulation, ensuring transferability to the real world. Experiments demonstrate the approach's effectiveness across various simulated scenarios, and we show real-world performance with systems composed of two and three Cassie robots. To best of our knowledge, this is the first scalable multi-biped payload transport system.

Learning Vision-Based Bipedal Locomotion for Challenging Terrain (Best paper finalist)

Helei Duan, Bikram Pandit, Mohitvishnu S. Gadde, Bart van Marum, Jeremy Dao, Chanho Kim, Alan Fern
IEEE International Conference on Robotics and Automation (ICRA), 2023

Reinforcement learning (RL) for bipedal locomotion has demonstrated robust gaits over moderate terrains using proprioceptive sensing, but these blind controllers fail in environments requiring visual perception. We propose a fully-learned system that enables bipedal robots to react to local terrain while maintaining travel speed and direction. Our approach trains a controller in simulation using a heightmap in the robot's local frame, and then uses data to train a heightmap predictor based on the history of depth images and robot states. With domain randomization, this approach achieves successful sim-to-real transfer without explicit pose estimation or fine-tuning. To best of our knowledge, this is the first example of sim-to-real learning for vision-based bipedal locomotion over challenging terrains.

Simple Masked Training Strategies Yield Control Policies That Are Robust to Sensor Failure

Skand Peri, Bikram Pandit, Chanho Kim, and Li Fuxin, Stefan Lee
Conference on Robot Learning (CoRL), 2024

Sensor failures are common in real-world robot deployments, leading to potential catastrophic outcomes like unexpected behaviors or damage. Previous approaches have tried to recover missing sensor data or adapt control policies during deployment. We propose training reinforcement learning (RL) policies that are robust to sensor failures using a multimodal encoder and a training strategy that simulates sensor failures by randomly dropping sensor modalities. Our evaluations on tasks such as bipedal locomotion and robotic manipulation, both in simulation and the real world, show that our method effectively handles sensor failures in various robot embodiments. The proposed method maintains training efficiency and sample efficiency, making it a promising solution for robust RL policies.

GRAD PROJECTS

Language for Model-Based Reinforcement learning

We applied multi-task learning (MTL) for robotic manipulation by using natural commands for controlling robotic arms. A seq2seq model architecture, incorporating autoencoders and attention models, was adopted to translate these natural commands into trajectories for the robotic arms. This approach demonstrated improved performance and generalization in robotic manipulation tasks through the integration of natural language processing and model-based reinforcement learning, compared to using model-based reinforcement learning alone.

Self-balancing platform using Reinforcement learning

I utilized a policy optimization approach to train a 2-DOF self-balancing platform within the physics-based simulator Mujoco. The training and inference pipeline was implemented using a combination of Mujoco, Gym, PyTorch, and TensorFlow. This approach led to successful simulation-to-reality transfer through TFLite model quantization, enabling low-power hardware inference. Additionally, the method demonstrated significant improvements compared to model-based approaches for balancing an object on a wobbling surface.

Multi-modal late-fusion architecture for learning Audio and Text

We developed a multi-modal neural network architecture that leverages state-of-the-art models such as Transformers, CNNs, and RNNs (LSTMs and GRUs) to analyze both the audio and textual aspects of songs. In this project, we incorporated and compared large language models (LLMs) like BERT, T5, and GPT, and utilized TensorFlow for training and fine-tuning the network to enhance its performance. This approach led to a significant improvement in accuracy, outperforming prior work by a substantial 34%, demonstrating the effectiveness of integrating multiple modalities.

Robust heuristic for Bipedal robot (Cassie) obstacle avoidance and point-goal navigation

Cassie is a bipedal robot popularly known for breaking Guinness world record for fastest 100m run. To achieve this feat, , the robot was trained in the simulator with high speed user command and proprioception as the input to the RL controller. However, such blind controller would fail in the situation where it has to react to the terrain, such as stepping on a block or avoiding an obstacle. Furthermore, the robot would constantly need the command from the user. In this project, I have built a autonomous heuristic for a robot to navigate in a terrain with obstacle with an addition of depth camera to the controller.

Machine common sense

In this particular task, I have developed the maneuver to pick a soccer ball that is surrounded by a lava region. The ball is unreachable because of its untraversable surrounding. However, there is a tool which a robot can use to push a ball out from untraversable region and then go pick up. I developed a heuristic in which a robot need to look around in a room with unknown map from its egocentric vision, find a ball, figure out the right tool and be able to use the tool correctly to pick up a ball. This work was sponsored by DARPA and our team from Oregon State University competed with team from other prestigious universities such as CMU, UCB, MIT and IBM where we stood a 2^nd place.

SOftware PROJECTS

Background Recorder

An android application that records audio / video, captures image even when the device screen is off.
Start recording anytime with volume button and capture anything in disguise.
It has 200K+ downloads, 4.1 ratings with over 500 daily active users.

Universal Translator

A translation service and platform to translate strings that is used in the software application and is packed with many more features.

Lab Reporting Software

In-browser Lab Reporting software. Progressive web app, installable on Mac, Windows, Linux Android and iPhone.

ShareByte

One place to write and share with everyone. An website where you can write anything and it synchronized over all browsers

OSU Waitlist count

Realtime minimalistic class wait-list counter for Oregon State University.
Use in this format https://osuclass.bikcrum.com/?crn=123,754

Live-Covid Tracker

An Android application showing COVID-19 statistics. This includes graph, time range selection and map view.

Chaincargo mobile apps

A mobile application for tracking the courier in delivery process.

Ekoplaza mobile apps

Ekoplaza is organic supermarket chain.

My cols

My cols has all the information of your favourite climbs, hills or cols worldwide. At the moment myCols has over 3800 cols and climbs.

LoveLock mobile apps

Tinder-like dating app (Android and iOS)

UNDERGRAD projects

Mix of programming, mechanics, and electronics.

Get In Touch

bikcrum@gmail.com

I’m Bikram,

~ teaching a machine to do stuff.

Education

EXPERIENCE

publication

GRAD PROJECTS

SOftware PROJECTS

opensource contributions

UNDERGRAD projects

Childhood ideas

Get In Touch