MMPUG: Multi-Model Perception Uber Good - The Robotics Institute Carnegie Mellon University
Home/MMPUG: Multi-Model Perception Uber Good
hero image
MMPUG: Multi-Model Perception Uber Good

This project seeks novel approaches to rapidly map, navigate, and search environments for situational awareness during time-sensitive combat operations. This project focuses on fast-moving autonomous vehicles designed for multi-modal exploration across various terrains and environments. To do this, we integrate and iterate upon previously designed algorithms for exploration, finding areas where common algorithms fail when used with high-speed UGVs and creating novel solutions to push the boundaries of what these vehicles are capable of. In addition, our system boasts a heterogeneous robot setup via legged robots that can aid exploration between multiple floors.

Fig: Heterogeneous agent system with three RC cars (UGVs) and two Spots (legged robots)

The current design combines a viewpoint-based exploration planner, trajectory libraries for rapid elimination of paths intersecting with obstacles, and a low-level planner capable of following paths at a high speed. An important aspect of the software architecture is its hierarchical design which allows the operator of the UGV to insert human feedback at any stage of the process. Specifically, this allows the operator to bias the exploration planner, give direct waypoint input, operate in a smart joystick mode (which follows the operator’s input while avoiding obstacles), or directly follow user input. The stack is then able to optimize the exploration process by utilizing different levels of autonomy to flexibly perform what the operator desires.

Fig: The third RC car in the convoy peels off to explore an unknown passage on the right

Our recent efforts involved creating a convoy of wheeled robots that could operate at higher speeds in both indoor and outdoor environments. We extended this capability to heterogeneous agents involving the Spot robots and illustrated that our optimal control algorithms are vehicle agnostic. Further, an explicit peel-off behavior was designed to ensure the trail vehicles could peel off from the convoy and act as radio beacons for the lead vehicles. In some scenarios, peeling off a part of the convoy to explore unknown terrains when there is a fork in the road has proven to be extremely helpful for heterogeneous multi-agent exploration. Once again, the hierarchical design allows the operator to provide feedback mid-exploration, and one of our new features includes bringing in new operators to handle peeled-off robots for a different mission.

Displaying 1 Publications


Share This Story!