Developing compact, interpretable latent representations for fine-grained robotic manipulation.
The goal is to learn parameterizations where a small number of latent dimensions can continuously
control policy scales (e.g., door-opening angle, motion speed) while preserving smoothness and
interpretability for downstream task adaptation.
- Built novel beta-VAE architecture with convolutional layers that preserves spatial and temporal structure of 6-DoF manipulation trajectories
- Engineered custom loss combining pairwise ranking and masked KL divergence to capture continuous policy scales
- Developed trajectory collection pipeline using MetaWorld simulation for model validation
- Created inverse kinematics visualization tool using Random Forest for real-time policy evaluation
- Pioneering LLM-assisted scale perception module to automate labeling and enable generalized policy learning
Improving multi-task performance of Latent Plan Transformers by integrating Model-Agnostic
Meta-Learning (MAML). The project addresses how agents can generalize across manipulation
tasks while producing stable, continuous control through learned strategy latents.
- Integrated MAML with Latent Plan Transformer, creating MPI-MAML (Meta-Planning as Inference)
- Rebuilt LPT codebase for D4RL Kitchen dataset compatibility, improving reproducibility across meta-learning benchmarks
- Engineered full MAML training loop with inner-loop K-shot adaptation and outer-loop meta-gradient accumulation
- Achieved 5-shot adaptation to novel manipulation tasks through rapid task generalization
- Currently implementing ANIL to test reduced inner loop computation for improved training efficiency
Developing intrinsically explainable AI architectures using biologically inspired neural field
representations and operator-theoretic tools. This work investigates how dynamical systems
theory can reveal the structure hidden inside learned dynamics.
Coherent Memory Structures in Neural Fields
- Developed framework using biologically inspired neural field representations for explainable AI
- Contributed to figures, boundary condition handling (periodic, Dirichlet, free-flow), and neuroscience literature review
- Paper under review at ICLR 2026
Conscious-Unconscious Neural Dynamics
- Applied Mori-Zwanzig projection operator formalism to separate neural dynamics into resolved (unconscious) and unresolved (conscious) states
- Achieved major MSE reduction on neural data (awake state: 3.140 → 0.101) compared to baseline DMD
- First-author paper under review at AAAI 2026 NeuroAI Workshop
HAVOK for Chaotic Time Series
- Clarified through mathematical derivation why HAVOK forcing emerges when truncated delay-embedded coordinates fail to span a Koopman-invariant subspace
- Validated on Lorenz and Rössler systems to localize nonlinear transitions during regime shifts
- Presenting at 2026 Joint Mathematics Meeting in Washington D.C.
Modeling and inferring agent goals within grid-based environments using Bayesian inference
and reinforcement learning. This project explored how observers can predict goals from
observed actions using structured probabilistic reasoning.
- Engineered foundational MDP solver using Value Iteration and Bellman updates to pre-compute optimal value functions for multiple hypothetical goals
- Developed gamma-discounted softmax policy extractor with temperature parameter for controlled action stochasticity
- Implemented robust Bayesian goal inference mechanism for real-time posterior updates based on observed actions
- Built full-stack, low-latency online visualization system using Python (RL backend) and Node.js/Socket.io (frontend) for real-time demonstration
Developed a lightweight attention-enhanced backbone for 6-DoF camera pose estimation,
achieving significant improvements in both translation and rotation error on the
King's College Cambridge dataset.
- Integrated CBAM (Convolutional Block Attention Module) with ResNet50 backbone
- Achieved 25% reduction in translation error and 35% reduction in rotation error
- Implemented learned-β loss for faster convergence and better generalization
- Trained on NVIDIA A100 GPU with 9,950 RGB frames