2024 Fall Robotics Colloquium: Manav Kulshrestha (Purdue University)

In this talk

1. 2:02

   Why robot decisions need to be legible

   Kulshrestha studies how people can infer a robot’s goals and beliefs from its behavior. He argues that understanding decisions matters especially in settings such as autonomous driving, military operations, and collaborative warehouses, where unclear behavior can add pressure or risk. He also notes that interpretability must be balanced with performance and social impact.

2. 13:13

   Learning reusable concepts

   Kulshrestha proposes representing scenes through meaningful components so robots can reason in ways people can follow. He uses vehicle recognition to show how features such as wheels and controls can help identify unfamiliar examples, and describes prior work that tracks face features to distinguish benign image changes from suspicious edits. These representations may help models generalize beyond their training examples and resist adversarial changes.

3. 20:20

   Planning block structures through dependencies

   For rearrangement tasks, Kulshrestha’s system segments RGB-D observations into object point clouds and predicts a directed graph of which blocks must support others. A topological ordering of that graph gives a build sequence, while disconnected substructures can be handled in parallel by multiple manipulators or a person. The team generates simulated structures in Bullet, checking stability by letting them settle, and reports generalization from two-level training examples to structures four or five levels deep.

4. 36:57

   Learning how agents influence one another

   Kulshrestha describes a multi-agent herding project in which controllable robots influence passive agents through repulsion. A learned forward model predicts the passive agents’ next states, while an inverse model estimates robot actions for a desired state; the approach shows preliminary results using linear interpolation for target paths. He points to possible uses in livestock herding, nanorobot guidance, and strategic drone operations.

5. 48:54

   Open directions: language and trust

   Kulshrestha considers bottleneck layers as a way to encourage models to form compact, potentially understandable representations, with regularization to align those representations with useful concepts. Future directions include using language knowledge to guide faster scene exploration and incorporating human trust into a robot’s objectives alongside task completion.