Lixin Xu

Lixin Xu

PhD student in Robotics

Posts by Tags

3D Gaussian Splatting

3D Point Cloud

3D Scene Generation

3D Vision

A100

AI

Bootstrapping and the Game of Entrepreneurship

11 minute read

Published:

Bootstrapping is the same pattern across compilers, AI, and startups: borrow external structure to cold-start, then recursively replace dependencies until the system sustains itself.

The Singularity is Near

1 minute read

Published:

Autonomous agents are now cheap and networked enough to spread faster than any institution can contain — the question is what order emerges after control becomes partial.

Intelligence Is Not Only Reasoning

9 minute read

Published:

Intelligence depends on retrieval and communication as much as reasoning — a powerful model limited by information gaps is still limited.

Cognitive Bandwidth in the AI Agent Era

8 minute read

Published:

As AI tools get more capable, the bottleneck shifts from tool friction to human cognitive bandwidth — compressing intent and steering effectively becomes the key skill.

AI Agents

The Singularity is Near

1 minute read

Published:

Autonomous agents are now cheap and networked enough to spread faster than any institution can contain — the question is what order emerges after control becomes partial.

Intelligence Is Not Only Reasoning

9 minute read

Published:

Intelligence depends on retrieval and communication as much as reasoning — a powerful model limited by information gaps is still limited.

Cognitive Bandwidth in the AI Agent Era

8 minute read

Published:

As AI tools get more capable, the bottleneck shifts from tool friction to human cognitive bandwidth — compressing intent and steering effectively becomes the key skill.

AI Ethics

AI for Science

Action Chunking

Action Coherence

Actuator Modeling

Admittance Control

Force Control and the Missing Layer in Embodied Intelligence

11 minute read

Published:

A late-night conversation over grilled skewers revealed a gap: most embodied AI researchers don’t think about what the robot arm is actually doing at the control level — and closing that gap might reshape how we think about action spaces.

Agentic AI

Alignment

Articulated Tools

Asynchronous Inference

Automated Discovery

Autonomous Research

Behavior Cloning

Benchmark

Benchmarking

Bimanual Interaction

Bimanual Manipulation

Blog Notes

Bootstrapping

Bootstrapping and the Game of Entrepreneurship

11 minute read

Published:

Bootstrapping is the same pattern across compilers, AI, and startups: borrow external structure to cold-start, then recursively replace dependencies until the system sustains itself.

CUDA

Claude Code

Claude Skills

CoRL

[Paper Notes] DayDreamer: World Models for Physical Robot Learning - CoRL 2022

1 minute read

Published:

Key information

  • This paper learns two models: a world model trained on off-policy sequences through supervised learning, and an actor-critic model to learn behaviors from trajectories predicted by the learned model.
  • The data collection and learning updates are decoupled, enabling fast training without waiting for the environment. A learner thread continuously trains the world model and actor-critic behavior, while an actor thread in parallel computes actions for environment interaction.

Code Generation

Codex

I Vibe Coded an Operating System

17 minute read

Published:

I built a scripting-first OS prototype in a weekend with Codex, where the system language gradually takes over its own environment from the host.

Coding Agents

Cognition

Cognitive Bandwidth in the AI Agent Era

8 minute read

Published:

As AI tools get more capable, the bottleneck shifts from tool friction to human cognitive bandwidth — compressing intent and steering effectively becomes the key skill.

Communication

Intelligence Is Not Only Reasoning

9 minute read

Published:

Intelligence depends on retrieval and communication as much as reasoning — a powerful model limited by information gaps is still limited.

Compliance Control

Computer Vision

Contact-Rich Manipulation

Coordinate Frames

Cross-Embodiment

Cross-Embodiment Learning

DINO

Data Augmentation

Data-Centric AI

Dataset

Deformable Manipulation

Depth Reconstruction

Dexterous Grasping

Dexterous Manipulation

Differentiable Simulation

Diffusion Models

Diffusion Policy

Digital Teleoperation

Digital Twins

DiscoRL

Discrete Diffusion

Distance Fields

Distillation

Distributed Training

Dynamic Manipulation

Dynamics Models

EMG

Economics

The Singularity is Near

1 minute read

Published:

Autonomous agents are now cheap and networked enough to spread faster than any institution can contain — the question is what order emerges after control becomes partial.

Egocentric Data

Egocentric Video

Egocentric Vision

Embodied AI

Embodied Agents

Embodied Intelligence

Force Control and the Missing Layer in Embodied Intelligence

11 minute read

Published:

A late-night conversation over grilled skewers revealed a gap: most embodied AI researchers don’t think about what the robot arm is actually doing at the control level — and closing that gap might reshape how we think about action spaces.

Embodied Planning

End Effectors

Engineering Notes

A Practical Map of LLM Training Ecosystems

8 minute read

Published:

A compact field note on how Megatron, Hugging Face, PyTorch, vLLM, SGLang, and verl fit together in a practical LLM training pipeline.

Entrepreneurship

Bootstrapping and the Game of Entrepreneurship

11 minute read

Published:

Bootstrapping is the same pattern across compilers, AI, and startups: borrow external structure to cold-start, then recursively replace dependencies until the system sustains itself.

Epistemology

Equivariance

Explainable AI

FSDP

Flow Matching

Force Control

Force Control and the Missing Layer in Embodied Intelligence

11 minute read

Published:

A late-night conversation over grilled skewers revealed a gap: most embodied AI researchers don’t think about what the robot arm is actually doing at the control level — and closing that gap might reshape how we think about action spaces.

Foundation Model

Foundation Models

Gaussian Splatting

Generalist Policy

Generative Models

Goal-Conditioned RL

Goal-conditioned

Grasp Synthesis

Grasping

Hand Morphology

Hand Motion

Hand-Object Interaction

Hugging Face

A Practical Map of LLM Training Ecosystems

8 minute read

Published:

A compact field note on how Megatron, Hugging Face, PyTorch, vLLM, SGLang, and verl fit together in a practical LLM training pipeline.

Human Data

Human Demonstration

Human Demonstrations

Human Priors

Human Video

Human Video Learning

Human Videos

Human-Object Interaction

Human-Robot Interaction

Human-Robot Transfer

Human-to-Robot Transfer

Humanoid

Humanoid Control

Humanoid Robotics

Humanoid Robots

Imitation Learning

Impedance Control

Force Control and the Missing Layer in Embodied Intelligence

11 minute read

Published:

A late-night conversation over grilled skewers revealed a gap: most embodied AI researchers don’t think about what the robot arm is actually doing at the control level — and closing that gap might reshape how we think about action spaces.

In-Hand Manipulation

Inference Acceleration

Information Retrieval

Intelligence Is Not Only Reasoning

9 minute read

Published:

Intelligence depends on retrieval and communication as much as reasoning — a powerful model limited by information gaps is still limited.

Information Theory

Instruction Following

Intelligence

Intelligence Is Not Only Reasoning

9 minute read

Published:

Intelligence depends on retrieval and communication as much as reasoning — a powerful model limited by information gaps is still limited.

Internet

The Singularity is Near

1 minute read

Published:

Autonomous agents are now cheap and networked enough to spread faster than any institution can contain — the question is what order emerges after control becomes partial.

Inverse Dynamics

Inverse Kinematics

Isaac Sim

KV Cache

Knowledge Transfer

LLM

LLM Agents

LLM Inference

LLM Training

A Practical Map of LLM Training Ecosystems

8 minute read

Published:

A compact field note on how Megatron, Hugging Face, PyTorch, vLLM, SGLang, and verl fit together in a practical LLM training pipeline.

LLM Workflows

Large Language Models

Latent Action

Latent Actions

Latent Dynamics

LeRobot

Legged Locomotion

Legged Robots

Loco-Manipulation

Long-Horizon Control

Long-Horizon Manipulation

Machine Learning Theory

Manipulation

Megatron

A Practical Map of LLM Training Ecosystems

8 minute read

Published:

A compact field note on how Megatron, Hugging Face, PyTorch, vLLM, SGLang, and verl fit together in a practical LLM training pipeline.

Memory

Memory Systems

Meta-Learning

Model-Based Control

Model-Based RL

Model-Free RL

Motion Capture

Motion Control

Motion Generation

Motion Imitation

Motion Retargeting

MuJoCo

Multi-Embodiment Learning

Multi-Object Grasping

Multimodal Learning

Multitask Learning

Novel View Synthesis

Offline RL

On-Policy Data

Online RL

Open Source

Operating Systems

Bootstrapping and the Game of Entrepreneurship

11 minute read

Published:

Bootstrapping is the same pattern across compilers, AI, and startups: borrow external structure to cold-start, then recursively replace dependencies until the system sustains itself.

I Vibe Coded an Operating System

17 minute read

Published:

I built a scripting-first OS prototype in a weekend with Codex, where the system language gradually takes over its own environment from the host.

Optimal Experiment Design

Optimization

PAL Robotics

PaliGemma

Paper Notes

Persona

Personal Thoughts

Force Control and the Missing Layer in Embodied Intelligence

11 minute read

Published:

A late-night conversation over grilled skewers revealed a gap: most embodied AI researchers don’t think about what the robot arm is actually doing at the control level — and closing that gap might reshape how we think about action spaces.

The Singularity is Near

1 minute read

Published:

Autonomous agents are now cheap and networked enough to spread faster than any institution can contain — the question is what order emerges after control becomes partial.

Intelligence Is Not Only Reasoning

9 minute read

Published:

Intelligence depends on retrieval and communication as much as reasoning — a powerful model limited by information gaps is still limited.

Cognitive Bandwidth in the AI Agent Era

8 minute read

Published:

As AI tools get more capable, the bottleneck shifts from tool friction to human cognitive bandwidth — compressing intent and steering effectively becomes the key skill.

Philosophy

Photometric Stereo

Physics

Physics-Based Animation

Pinocchio

Planning

Point Clouds

Policy Improvement

Policy Learning

Post-Training

Pretraining

Productivity

Cognitive Bandwidth in the AI Agent Era

8 minute read

Published:

As AI tools get more capable, the bottleneck shifts from tool friction to human cognitive bandwidth — compressing intent and steering effectively becomes the key skill.

Programming Languages

I Vibe Coded an Operating System

17 minute read

Published:

I built a scripting-first OS prototype in a weekend with Codex, where the system language gradually takes over its own environment from the host.

Project Notes

I Vibe Coded an Operating System

17 minute read

Published:

I built a scripting-first OS prototype in a weekend with Codex, where the system language gradually takes over its own environment from the host.

Prompt Engineering

PyTorch

A Practical Map of LLM Training Ecosystems

8 minute read

Published:

A compact field note on how Megatron, Hugging Face, PyTorch, vLLM, SGLang, and verl fit together in a practical LLM training pipeline.

RGB Perception

RL Algorithms

RSS

Real-Time Control

Real-World RL

Reference Tracking

Reinforcement Learning

A Practical Map of LLM Training Ecosystems

8 minute read

Published:

A compact field note on how Megatron, Hugging Face, PyTorch, vLLM, SGLang, and verl fit together in a practical LLM training pipeline.

[Paper Notes] DayDreamer: World Models for Physical Robot Learning - CoRL 2022

1 minute read

Published:

Key information

  • This paper learns two models: a world model trained on off-policy sequences through supervised learning, and an actor-critic model to learn behaviors from trajectories predicted by the learned model.
  • The data collection and learning updates are decoupled, enabling fast training without waiting for the environment. A learner thread continuously trains the world model and actor-critic behavior, while an actor thread in parallel computes actions for environment interaction.

Representation Learning

Research Automation

Research Notes

Retargeting

Reward Model

Reward Modeling

Rigid Dynamics

Robot Calibration

Robot Co-Design

Robot Dynamics

Robot Foundation Models

Robot Learning

Robot Manipulation

Robotic Manipulation

Robotics

Force Control and the Missing Layer in Embodied Intelligence

11 minute read

Published:

A late-night conversation over grilled skewers revealed a gap: most embodied AI researchers don’t think about what the robot arm is actually doing at the control level — and closing that gap might reshape how we think about action spaces.

[Paper Notes] DayDreamer: World Models for Physical Robot Learning - CoRL 2022

1 minute read

Published:

Key information

  • This paper learns two models: a world model trained on off-policy sequences through supervised learning, and an actor-critic model to learn behaviors from trajectories predicted by the learned model.
  • The data collection and learning updates are decoupled, enabling fast training without waiting for the environment. A learner thread continuously trains the world model and actor-critic behavior, while an actor thread in parallel computes actions for environment interaction.

Robustness

Runtime Systems

I Vibe Coded an Operating System

17 minute read

Published:

I built a scripting-first OS prototype in a weekend with Codex, where the system language gradually takes over its own environment from the host.

SGLang

A Practical Map of LLM Training Ecosystems

8 minute read

Published:

A compact field note on how Megatron, Hugging Face, PyTorch, vLLM, SGLang, and verl fit together in a practical LLM training pipeline.

Sample Efficiency

Scaling Laws

Science Robotics

Scientific Discovery

Self-Hosting

Bootstrapping and the Game of Entrepreneurship

11 minute read

Published:

Bootstrapping is the same pattern across compilers, AI, and startups: borrow external structure to cold-start, then recursively replace dependencies until the system sustains itself.

Sensor Design

ShadowHand

Sim-to-Real

Sim2Real

Simulation

Single-Image Reconstruction

Skill Chaining

Spatial Grounding

Spatial Reasoning

State Representation

Symbolic Regression

Synthetic Data

System Identification

Systems

Tabletop Scenes

Tacit Knowledge

Tactile Sensing

Tactile Simulation

Task and Motion Planning

Technical Notes

Teleoperation

Test-time Guidance

Theory

Tokenization

Tokenizer

Trajectory Optimization

Transformers

URDF

Unsupervised RL

VLA

VLM Fine-tuning

Value Functions

Video Diffusion

Video Foundation Models

Video Generation

Video World Models

Vision Language Action

Vision-Based Tactile Sensors

Vision-Language Models

Vision-Language-Action

Vision-Language-Action Models

Vision-Tactile Learning

Visuo-Tactile

Visuotactile Simulation

Whole-Body Control

Whole-Body Manipulation

World Action Models

World Model

World Models

Zero-shot Generalization

Zero-shot Manipulation

update

Updating website

less than 1 minute read

Published:

A major site update done mostly by Codex — page refactoring, content restructuring, and bilingual support.

vLLM

A Practical Map of LLM Training Ecosystems

8 minute read

Published:

A compact field note on how Megatron, Hugging Face, PyTorch, vLLM, SGLang, and verl fit together in a practical LLM training pipeline.

website

Updating website

less than 1 minute read

Published:

A major site update done mostly by Codex — page refactoring, content restructuring, and bilingual support.