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.

Led benchmarking of self-correction in LLMs, evaluating performance across Llama 3.2 1B, Llama 3.1 8B, and Mathstral 7B on MATH dataset. Achieved accuracy rates of up to 41.8%, identifying high Correct-to-Incorrect rates (46.9%) and low Incorrect-to-Correct rate improvements (3.78%). Engineered a multi-turn reinforcement learning framework (SCoRe) for fine-tuning.

Re-implemented the LIPINC-V2 Vision Temporal Transformer, a SOTA architecture for detecting deepfakes by analyzing spatio-temporal video. I first validated my from-scratch model by reproducing the paper's 0.98 AP results on the LipSyncTIMIT benchmark. To expand the project, I then engineered a custom data generation pipeline, creating a new 2,300-sample dataset. On this new dataset, I established a strong CNN-LSTM baseline model, achieving 94.9% accuracy.

Developed an onset detection technique to isolate the Percussion Solo section in concert audio, applying clustering algorithms (K-means, DBSCAN) using audio processing libraries.