[Paper Notes] WARP: Whole-Body Retargeting for Learning from Offline Human Demonstrations
Published:
Blog posts
Browse all posts, or filter them into personal thoughts and paper notes.
2026
Why On-Policy Data Matters in Post-Training
Published:
[Paper Notes] Human Universal Grasping
Published:
Before the Robot GPT-3.5 Moment: Latent Actions, World Models, and Embodied Memory
Published:
A reflective essay on latent actions, world models, KV-cache memory, and why robotics may still be before its GPT-3.5 moment.
A Practical Map of LLM Training Ecosystems
Published:
A compact field note on how Megatron, Hugging Face, PyTorch, vLLM, SGLang, and verl fit together in a practical LLM training pipeline.
Force Control and the Missing Layer in Embodied Intelligence
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.
Bootstrapping and the Game of Entrepreneurship
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.
Michael Polanyi: Tacit Knowledge, Personal Knowing, and What AI Still Cannot Tell
Published:
Polanyi’s framework of tacit knowledge and personal knowing offers a lens for understanding what AI still cannot do — and why embodied, context-dependent skill remains hard to formalize.
Distilling Persons into Agents: A Survey of Recent ‘Person-as-Skill’ Projects
Published:
A survey of projects that compress human personas into AI agent skills — from departed colleagues to public figures — and what this trend reveals about AI, memory, and identity.
[Project Notes] Robotic World Model Lite
Published:
[Project Notes] PAL Robotics in mjlab
Published:
The Singularity is Near
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.
How to Align Teleoperation Devices with Robot End Effectors
Published:
The hardest part of teleoperation is aligning coordinate frames between input devices and robot end effectors — a practical walkthrough.
Inverse Kinematics for Any URDF: Tasks, Weights, and Solver Design
Published:
A practical guide to formulating IK as weighted nonlinear least-squares, with task definition, damping, and solver design that works for any URDF.
I Vibe Coded an Operating System
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.
Intelligence Is Not Only Reasoning
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
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.
Updating website
Published:
A major site update done mostly by Codex — page refactoring, content restructuring, and bilingual support.
2024
[Paper Notes] SAM-RL: Sensing-Aware Model-Based Reinforcement Learning via Differentiable Physics-Based Simulation and Rendering - RSS 2023
Published:
Key information
- Model-based reinforcement learning (MBRL) is potentially more sample-efficient than model-free RL.
- Integration of differentiable physics-based simulation and rendering facilitates the learning process.
- Pipeline: Real2Sim -> Learn@Sim -> Sim2Real to produce efficient policies.
[Paper Notes] DayDreamer: World Models for Physical Robot Learning - CoRL 2022
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.