Sitemap

Coded on-policy and off-policy reinforcement learning algorithms from scratch for robot navigation by following a wall using LiDAR data. Implemented in ROS and Python, using Gazebo for simulation.

Designed a bidirectional communication system to resolve ambiguous fetch and return tasks. Used VLMs to determine whether a request is ambiguous based on the current scene and request more information if it is. Implemented object detection and mobile pick and place on a Stretch 3 mobile robot.

Evaluated our pipeline with tasks related to grabbing a drink as well as more general tasks that had a variety of objects. Clarification was correctly requested 100% of the time for the drink tasks but only 20% of the time on general tasks. This demonstrates the potential for out-of-the-box VLMs to clarify requests in simple tasks, but this can fail in more complex and ambiguous scenarios.

Novel approach for collecting data to fine-tune imitation learning policies using blended shared control during interventions for better autonomous performance. In comparison to full human control during interventions (HG-DAgger), our approach results in:

• Higher autonomous performance by over 30 percentage points on two long horizon real-world tasks
• ~15% faster data collection
• 40% smoother transitions from autonomous rollouts to human interventions and 63% smoother transitions from interventions to rollouts

Implemented and compared on-policy (PPO), off-policy (SAC), and model-based (MBPO) on object grasping in the RoboSuite simulation. Compared policy performance on vision-based and state-based inputs and with warmstarting from imitation learning. See more details here.