Design study for a low-cost, minimal exoskeleton glove for capturing human hand data at factory scale — active motion capture handling the knuckle's 2 DoF, with finger topology referenced from Naked Prosthetics and ART Glove, and a separate tactile knitwear layer (Paxini / AgileTact / piezoresistive FPCB sheets). Includes my FreeCAD sources, comparison reports and BOMs; the full ~6.5 GB research archive (reference glove/hand repos, CAD, papers, videos) is on HuggingFace.

Solution to the Intrinsic robotics challenge: inserting a cable plug into randomized ports on a task board with a UR5e arm, using only RGB cameras and wrist force/torque feedback — no ground truth in the control path. Combines YOLO-based port detection with CAD-plane back-projection and module gating (58mm → 2mm error), marker-anchored board pose estimation (<1.3° yaw), singularity-avoiding joint-space motion planning, and sign-insensitive contact detection for compliant insertion.

Multiview hand-triangulation pipeline reproducing "What Matters When Cotraining Robot Manipulation Policies on Everyday Human Videos?". Lifts EgoExo4D multi-view RGB (one egocentric Aria fisheye + four exocentric GoPros) to 3D hand keypoints via YOLO detection, WiLoR 2D keypoints, and DLT + IRLS triangulation — reaching ~24 mm 3D MPJPE against ground truth, roughly 8× better than monocular estimates.

LoRA adapter for exocentric-to-egocentric video generation, reproducing the EgoX paper on a single 24GB GPU instead of the paper's 4× H200. To fit inference on my own GPU I downscaled the video resolution and trained a rank-128 LoRA on top of the Wan2.1-I2V-14B diffusion model with Geometry-Guided Self-Attention, NF4 4-bit quantization, and 8-bit AdamW, on Ego-Exo4D.

Minimal reimplementation of Sensor2Sensor: Cross-Embodiment Sensor Conversion for autonomous driving, trained on a single RTX 3060. Generates LiDAR point clouds from surround camera images via a DINOv3-conditioned latent diffusion model over a compact LiDAR range-image VAE, on nuScenes (held-out Chamfer distance ~2.0 m for the diffusion U-Net).

Simulation-based reproduction of ENPIRE's robot-policy self-improvement loop (Environment → Policy improvement → Rollout → Evaluation) on the Push-T task. Multiple Claude coding agents iteratively improve control policies in git-isolated worktrees and share recipes across a fleet — collaborative agents reached ~50% held-out success vs. 14.6% for isolated agents.

Tennis match analysis pipeline with ball tracking (TrackNet/TrackNetV4/WASB-SBDT), court detection, player pose estimation, and racket detection from broadcast video.

3D pose reconstruction of tennis players from monocular broadcast video. Lifts 2D keypoints to a 3D skeleton for biomechanics analysis.

An AI-powered job search and auto-application tool with a simple UI for browsing and applying to roles.

Online vector quantization with near-optimal distortion rate. 2-bit KV cache quantization (~8x memory reduction) for HuggingFace transformer inference, within ~2.7x of the Shannon limit.

Implementation of the Action100M video processing pipeline for generating hierarchical video action annotations. Uses V-JEPA 2 for temporal segmentation and Qwen3-VL for caption generation.

Autonomous agent framework that implements research papers from PDF to working code. 6-phase pipeline: paper extraction, contract definition, data proof, component implementation, integration, and benchmarking.

Recreation of the 3D reconstruction pipeline from the LingBot world model paper. Generates detailed 3D point clouds from video sequences using VGGT.

Digital twin of Unitree G1 humanoid robot in a 3D Gaussian Splatting scene. Browser-based split-screen viewer with keyboard control for humanoid locomotion in photorealistic environments.

Implementation of Copilot4D: Learning Unsupervised World Models for Autonomous Driving via Discrete Diffusion. A world model approach for self-driving vehicles.

Diffusion-based trajectory prediction using LiDAR data. Trained on KITTI dataset with encoder and diffusion model for autonomous driving trajectory forecasting.

Implementation of DreamerV4: Training Agents Inside of Scalable World Models. World model architecture for reinforcement learning with video tokenizer and dynamics prediction.

Implementation of Humanoid World Models: Open World Foundation Models for Humanoid Robotics. Training world models for humanoid robot control and simulation.

Implementation of Genie world model - a generative video model that learns latent actions from unlabeled video data. Includes video tokenizer, latent action model, and dynamics model.

Robotic piano playing simulation in NVIDIA Isaac Sim with dexterous hands. Features MIDI playback, collision handling, and reinforcement learning for piano performance.

Reinforcement learning for SO101 robotic arm in NVIDIA Isaac Lab. Custom SO101 stacking cubes task created from scratch with GPU-accelerated parallel environments for training manipulation policies.

Projects and exercises for learning ROS2 Navigation Stack (Nav2) using TurtleBot3 in Gazebo simulation. Covers SLAM, autonomous navigation, and the complete Nav2 stack implementation.

Fine-tune the model on YouTube videos that show first-person hand movements, extract motion data, refine it in Blender, and retarget it for specific hand URDF models.

Degrees of freedom analysis and comparison of the Berkeley Humanoid platform. Cost estimated at $4.9K.

A collection of gymnasium environments for Human-In-the-Loop (HIL) reinforcement learning, compatible with Hugging Face's LeRobot codebase.

The LeRobot framework along with the SVLA SO101 pick-and-place dataset for robot learning.

Fork of Physical Intelligence's OpenPI with flow matching for robot policy learning. Vision-language-action model with continuous action generation using conditional flow matching for high-frequency control.

Tools for converting MMD (MikuMikuDance) motions to robot armature format for Isaac Lab RL training. Includes motion extraction from Blender, inverse kinematics retargeting to robots like Unitree G1/H1.

Expressive robotic hand design. More dexterous configurations using cables and deported actuators in the forearm.

Code and environments for reinforcement learning of drone swarms using Unreal Engine 4 and Microsoft AirSim. Includes multi-agent RL implementations.

Jewel Thief Circuit simulation found in most electric mosquito zappers.

HumanoidBench: a simulated humanoid robot benchmark with 15 whole-body manipulation and 12 locomotion tasks, with code for environments and training.

Gym environments for humanoid robot locomotion research.

Isaac Launchable offers a simplified approach to installing and using Isaac Lab and Isaac Sim.