Fall 2022 GRASP SFI - Srinath Sridhar, Brown University

In this talk

1. 2:00

   Why robots need 3D object understanding

   Sridhar connects everyday tasks such as inspecting produce, construction, and precise cutting to the perception skills robots need. He focuses on recognizing object properties, including shape, articulation, position, and orientation, as a foundation for manipulation.

2. 9:02

   ShapeCrafter for iterative shape design

   ShapeCrafter uses language prompts to generate and revise 3D shapes, keeping a consistent sample as edits accumulate. To support this process, the team expanded Text2Shape descriptions into linked phrase and shape examples, and used a quantized implicit representation to retain local details. The demonstrated model covers chairs and tables; Sridhar notes that details can be imperfect and edits may lose track of earlier shape features.

3. 34:59

   CONDOR for canonical object orientation

   CONDOR estimates consistent orientations for full or partial point clouds without requiring labeled canonical poses. Its design uses features equivariant to point ordering and rigid transformations, then derives a canonical direction and encourages consistency across object instances. The method can also produce consistent part segmentations, but symmetry, noise, and clutter remain difficult cases.

4. 50:05

   Extending the work to motion and real data

   Sridhar describes ongoing work on articulated objects, including discovering parts, joints, and joint limits, and on finding canonical arrangements for groups of shapes. His lab is building a capture rig with 54 synchronized cameras recording at 120 frames per second to study hands, object interactions, and small robots. He also points to a neural fields paper collection with more than 400 entries.