Projects | Ming Lin

Unified Humanoid Whole-Body Control

Fri, 06 Nov 2026 00:00:00 +0000

We are researching a unified whole-body control model that can generalize across humanoid robot platforms. The model accepts a robot’s hardware configuration as context, with the long-term goal of operating across embodiments without platform-specific fine-tuning.

NVIDIA GR00T is the current base model, although the method is designed to extend to other control architectures. Initial cross-platform generalization has been demonstrated, and work continues on robustness and transfer.

Synthetic-Data RL for VLA Models

Fri, 06 Nov 2026 00:00:00 +0000

We are building a framework for training vision-language-action models to perform complex manipulation tasks in data center environments without requiring human demonstrations or teleoperation data.

Pre-programmed policies initialize reinforcement learning, while diverse synthetic scenes provide scalable experience. The goal is to reduce the cost and operational burden of collecting task-specific demonstration data.

Robot-Agnostic Navigation

Fri, 06 Nov 2026 00:00:00 +0000

We are developing a lightweight navigation runtime for a broad range of robotic platforms. The system uses a graph-based representation and deep-learning-derived visual features as anchor points, minimizing robot-specific dependencies and simplifying deployment across different robot embodiments.

The robot-agnostic architecture is implemented, with broader cross-platform validation in progress.

Multi-Robot Mission Orchestration

Fri, 06 Nov 2026 00:00:00 +0000

We are developing a mission-orchestration system for multiple robots executing concurrent tasks in data centers. Robot availability, task priorities, resource constraints, and operating conditions change continuously, creating a dynamic, non-convex scheduling problem.

Our approach uses deep reinforcement learning to optimize mission allocation and scheduling across the fleet.

Efficient Model Design

Fri, 06 Nov 2026 00:00:00 +0000

My work in neural architecture search focuses on training-free search, hardware-aware model design, and efficient architectures for vision and language models.

Tiny-NAS is a distributed, full-stack framework for zero-shot neural architecture search, with hardware-latency prediction for GPU and mobile deployment. Zen-NAS is a fast, training-free method for finding models that balance predictive accuracy and inference efficiency.