Find a thread.

The quantitative language for learning, estimation, geometry, and uncertainty.

The software abstractions that make experiments reproducible and systems maintainable.

Algorithms, systems, networking, and concurrency beneath intelligent machines.

Real-time hardware, sensors, and actuators where computation meets the physical world.

Learning useful structure from data, objectives, and feedback.

Differentiable representations optimized from data and objectives.

Turning images and video into structured state, identity, and affordances.

Language models as predictors, representations, and interfaces for tools.

Aligning visual and language representations so models can ground symbols in scenes.

Policies learned through rewards, interaction, and long-horizon consequences.

Models, tools, memory, and feedback loops organized around goals.

The physical models and feedback laws that make motion predictable and controllable.

Sensors and estimators that turn imperfect measurements into usable state.

Localization, mapping, planning, and control coordinated under motion constraints.

Middleware, communication, and runtime structure for composing robot systems.

Contact, grasp, and motion planning for robots that handle the physical world.

Balance, locomotion, and whole-body motion under contact and embodiment.

Simulation environments and digital twins for testing learning and control safely.

Perception that is situated, active, and useful for choosing what to do next.

Wayfinding and exploration grounded in language, geometry, and action.

Long-horizon, contact-rich manipulation organized around affordances and feedback.

Whole-body policies and contact-aware skills for humanoid agents.

Vision-language-action models that connect perception, language, and robot policy.

Predictive state, dynamics, and memory for imagining what an action may cause.

Agents that plan, remember, recover, and act over extended physical tasks.

Policies learned from demonstrations, rewards, and repeated interaction with a robot.

Demonstrations, synthetic worlds, and transfer procedures that make robot learning scalable.

用一个最小的计算图，把链式法则、自动微分和实验记录放在同一张工作台上。

A complete ROS 2 navigation stack for an Ackermann vehicle, connecting state estimation, planning, control, and safety on a real competition platform.

A MOTR-oriented research system that separates geometry/motion, appearance identity, and existence/lifecycle into explicit state domains.

A Temporal Shift Module pipeline for real-time gesture recognition, with data preparation, training configuration, checkpoint loading, and webcam inference.

Reading note in progress.

Reading note in progress.

Reading note in progress.

Reading note in progress.

Reading note in progress.

Reading note in progress.