Anuva Banwasi

Developing a persistent, object-centric memory system linking RGB-D perception with adaptive task planning for mobile manipulators. Built a real-time scene-graph that synchronizes perception, action, and memory, updating as the robot explores and interacts with its environment. Leveraged VLMs grounded with physical awareness to predict object affordances and choose appropriate motion primitives for interaction. Testing in Mujoco/IsaacSim and real-world experiments with TidyBot.

Previously built a dexterous grasp evaluation pipeline using 3D Gaussian Splatting. Designed custom VQ-VAE architecture to compress 3DGS into latent codes and trained a multi-headed evaluator predicting grasp success, collisions, and stability, achieving 95.1% grasp success rate in simulation, surpassing BPS/NeRF baselines.

Built an emotion-aware conversational AI system combining ASR, prosodic features, and LLM reasoning to detect sudden emotion shifts from neutral or happy to sad or anger. Multimodal fusion of transcripts, audio embeddings, and dialogue context. Trained on 150+ IEMOCAP videos (~10GB), improving accuracy from ~40% to 78%.

Led a team of 4 researchers to build zero-shot learning framework that assesses step-by-step task execution for long-horizon tasks. Algorithm describes predicted step outcome with LLM, evaluates video data against language outcome description with VLM, and assesses incremental progress with custom scoring metric. Achieved ~95% accuracy across diverse household and kitchen tasks. First-author publication at International Conference for Robotics and Computer Vision (ICRCV) 2023.

Built vector autoregression models to forecast EV adoption using policy incentives and charging infrastructure data. Produced a first-author publication at ITISE 2024.

Designed a NeRF-inspired model that learns robot kinematics from single-view video by predicting occupancy from 3D coordinates and joint angles. Used PyBullet-derived camera intrinsics to reconstruct 3D geometry from 2D single-view data.