GO1 // CAIRO LAB
Go1 research portfolio • two related studies

Go1 Autonomous Navigation Research

From learning motion through imitation to studying social navigation decisions with VLMs.

This site documents two related Go1 navigation studies. The first asks whether Go1 can learn real-time navigation behavior from demonstration data. The second asks how a robot that can already move should make higher-level decisions around people.

Imitation Learning Motion learning VLM Social Navigation Social decision study
Research Snapshot
0
Related Studies
motion + social decision studies
0
Curated Social-Nav Bags
curated benchmark bags
0
VLM Decision Labels
stop, forward, left, right, review
GO1 SYSTEM VIEW
Motion Social Decision Safety
Fast controller + semantic reasoner

Go1 Research Map

The portfolio makes the most sense as three connected layers: motion, social decision, and the boundary between them.

IMITATION LEARNING NAVIGATION STUDY 1
Motion Layer

The earlier study focuses on real-time learned motion behavior on Go1.

Question How does Go1 learn to move?
Inputs Front-camera imagery and demonstration data
Outputs Real-time motion behavior
What was real Deployed controller on Go1
Key Signals
Traversability Obstacle Risk Stability
focus low-level motion and reactive safety
question whether a learned controller can run online on the robot
next see how the project changed once people entered the scene
Imitation Learning →

Why the Project Split

The second study did not replace the first one. It came from a limit in what the first system could represent once people entered the scene.

Reliable movement first

  • A robot first needs a controller that can run online without becoming unstable.
  • That earlier work focused on reactive motion, smoothing, and deployment on Go1.

Representation gap second

  • Binary stop/go behavior was too coarse for crossings, yielding, and ambiguity.
  • The problem became not only control quality, but decision representation.

Final system view

  • Fast controller handles low-level motion and local safety.
  • Slower VLM layer contributes semantic guidance through safety projection.

Where to Read Next

The overview is only the map. The technical details, limitations, and observations live on the two study pages.

Study 1: Imitation Learning Navigation

  • Real-time learned controller on Go1
  • Low-level signals: traversability, obstacle risk, stability
  • What reactive control did and did not capture
Imitation Learning → Research Overview →

Study 2: VLM Social Navigation

  • Reported benchmark over curated Go1 social-navigation bags
  • Controller-side integration studies around high-level semantic decisions
  • Safety-projected use instead of raw VLM motor control
VLM Social Navigation → Social Navigation Overview →

Research Notebook

Technical notes, model details, benchmark setup, and deployment boundaries behind both Go1 studies.

Open Notebook →