News

Publications

TL;DR: FALCON enables various heavy-duty humanoid loco-manipulation tasks via a new dual-agent force-adaptive RL framework.

Industry Experience

Built an autonomy system that fuses 6D pose estimation (NVIDIA FoundationPose) with motion-capture localization for precise perception and navigation. This allowed the Unitree G1 to autonomously manipulate totes and operate effectively in real-world factory workflows.

Benchmarked self-correction capabilities across Llama 3.2 1B, Llama 3.1 8B, and Mathstral 7B on the MATH dataset, achieving 41.8% accuracy. Implemented the proposed multi-turn reinforcement learning framework (SCoRe) for fine-tuning LLMs to improve self-correction.

Built a Vision-Language Navigation (VLN) system that answered natural language queries by combining Gemini 2.5 Pro embodied reasoning with a custom ROS state machine. The system produced numerical answers, object references, or waypoint plans under a strict 10-minute limit.

Developed a full-stack motion tracking and imitation learning system enabling a Franka Emika Panda robot to dynamically track and mimic human sword motion. Integrated YOLOv8 for real-time object segmentation with HSV thresholding for pose detection, transforming sword trajectories into the world frame to continuously update the robot's pose for motion imitation.

Re-implemented the LIPINC-V2 Vision Temporal Transformer for deepfake detection, validating it by reproducing the paper's 0.98 AP on LipSyncTIMIT. Engineered a custom 2,300-sample dataset and established a CNN-LSTM baseline achieving 94.9% accuracy, revealing critical insights into dataset bias and model generalization.

© 2025 Ishita Gupta · Template from Jon Barron