Research Blog
研究日誌・論文解説・実験結果を、新しい順に掲載しています。
Do Robots Really Need World Models? CoRL 2026 Takes On World Models vs. End-to-End VLA Head-On
Notes on the CoRL 2026 workshop "Do Robots Need World Models?", the debate between explicit world models and end-to-end VLA policies, and why the more useful question may be where a world model should live in a robot system.
Reinforcement Learning That Gives VLAs "Reflexes": Real-Time EXPO-FT and Its Application to Vision-Tactile World Models
A reading of "Reinforcement Learning for Real-Time Vision-Language-Action Policies" (Real-Time EXPO-FT), which pairs a large VLA with a small RL-trained edit policy and Q-function, and ideas for combining it with a Vision-Tactile JEPA world model for failure-aware manipulation.
ポートフォリオサイトを公開しました
このサイトの構成と、今後 Research Blog に載せていく内容についてのメモです。
Rethinking Next-Generation VLA from MolmoAct 2: Combining Tactile Sensing and World Models
A comparison between Allen Institute for AI's MolmoAct 2 robot foundation model and the author's master's research on a Vision-Tactile JEPA World Model for long-horizon VLA, exploring how tactile sensing and future-state prediction could help robots recover from failure.
D2: From Drawing Diagrams to Managing Them as Code — Diagram as Code for the AI Era
Notes on D2, a Diagram-as-Code tool, comparing it with Mermaid and draw.io MCP, and looking at why managing architecture diagrams as text pairs especially well with AI Coding Agent-driven development.
Looking into Marigold V2: Depth Estimation with a Diffusion Transformer
Notes on Marigold V2, a monocular depth estimation model from HUAWEI Bayer Lab, EPFL, and the University of Bologna that repurposes an image-generation Diffusion Transformer for single-step dense prediction, and what it suggests about reusing generative foundation models to understand the world.
Starting Robot Arm Teleoperation
A short note on beginning hands-on teleoperation of a robot arm as part of the graduate research theme, working through the gap between theory and practice by trial and error.
From rg to zg: Local Search for the AI Agent Era
Notes on zg (zvec-grep), an open-source local code search tool from Alibaba Cloud that combines vector search, BM25, and rg-style exact match, and why it matters for AI Coding Agents.
Starting to Use Grok Bot
A short note on trying out Grok Bot on the Cursor Pro plan, and the novelty of it assigning each user their own VM.
Gaussian Splatting: From "Seeing" 3D to a "Touchable" World
Notes on World Labs' Splat World case study, where Gaussian Splatting is turned into a physically reactive VR material rather than a static background, and what that could mean for robot simulation and digital twins.
Migrating from Tmux to Herdy for Managing AI Coding Agents
A short note on switching from a Ghostty + Tmux setup to Herdy in order to manage multiple AI Coding Agents more efficiently.
Collecting Robot Learning Data Without Carrying a Robot: Reading the UMI Paper
Notes on Universal Manipulation Interface (UMI), a hand-held data-collection device that narrows the human-to-robot embodiment gap through hardware design, and how it relates to the author's own Vision-Tactile Latent World Model research for failure-aware VLA.
Does a VLA Need to Understand Its Own Body? Thinking Through Embodiment via Whole-Body VLA, Tactile Sensing, and World Models
An exploratory essay asking whether Vision-Language-Action (VLA) models actually understand a robot's own body, distinguishing whole-body control from embodiment, and proposing an 'Embodied Latent State' and 'Body Model' as a research framing.
Claude Code Cross-Session Messaging
A quick reaction to the Claude Code cross-session messaging feature, and its potential for sending messages to sessions on other machines.
A Diagram of My Research Environment
A rough diagram sharing the environment used for research, both at grad school and at work.
VLA-JEPA Has Arrived
Notes on VLA-JEPA, which brings the JEPA approach into VLA pretraining to address weaknesses in physical-law understanding and long-horizon tasks, with a comparison against V-JEPA2.
On Loop Engineering
A short note on the progression from prompt engineering to context engineering, harness engineering, and now loop engineering in AI Coding Agent workflows, with pointers to further reading and the Agent SDK.