Study 1 • low-level navigation and deployment

Imitation Learning Navigation

Learning motion from demonstrations.

Research question: can a quadruped robot learn useful navigation behavior from demonstration data instead of relying entirely on hand-designed rules? This study focuses on low-level motion behavior: moving forward, slowing or stopping, reacting to obstacles, and staying stable enough to use on hardware.

System Design Back to Overview

Core Reactive Signals

Robot Deployment Study

Real-Time Controller Path

PHASE 1 TELEMETRY GO1

Phase 1: Real-Time Reactive Navigation Metrics

These metrics describe whether the robot is moving safely and reliably enough to use. They mattered in the earlier navigation system, but they were never the whole story once people entered the scene.

Traversability 72%

Obstacle Risk 18%

Stability 88%

signals reactive motion and safety cues

runtime smoothing, clipping, and conservative overrides

limit little explicit representation of human intent

Research Motivation

The earlier problem was mostly about whether a learned navigation pipeline could actually run on the robot without becoming unstable or unusable.

What made this hard

Manual navigation rules are difficult to tune across different spaces and lighting conditions.
Real robots have latency, sensing noise, and imperfect control response.
A model that looks reasonable offline can still fail if the online loop is too noisy or too abrupt.

Why imitation learning

Teleoperated demonstrations provide motion examples directly from the robot's own sensing context.
The learned policy can absorb corridor following, turning, and stop/go behavior without hand-writing every decision rule.
The practical question was not only model accuracy. It was whether the whole loop could stay responsive and stable on Go1.

System Design

This phase centers on a compact visual policy that could be integrated into a ROS runtime and kept light enough for real-time use.

Data and Runtime

Go1 teleoperation and rosbag recording
Front-camera imagery aligned with motion-command supervision
ROS pipeline for online inference on the robot

Model Family

ResNet-18 visual backbone
Short temporal history or GRU-style context where useful
Velocity-style outputs and stop-gate style filtering where supported

Deployment Constraints

Inference had to remain lightweight enough for online use.
Smoothing and clipping mattered because raw predictions are not automatically usable on hardware.
The controller had to be conservative enough to stay stable under noise and timing variation.

Metrics and Signals

Traversability, obstacle risk, and stability should be read as low-level motion and safety signals from the earlier navigation system, not as a complete account of navigation.

What these signals were good for

Traversability: whether local space looked passable enough to keep moving.
Obstacle risk: whether the current path looked likely to require slowing or stopping.
Stability: whether the control loop stayed usable enough for real robot motion.

What they did not say

They do not explain whether a person is crossing, receding, or entering the frame late.
They do not provide a direct way to represent yielding left, yielding right, or deferring under uncertainty.
That gap became the main reason the project eventually needed a second study.

What This Phase Achieved

The contribution of this phase was not that it solved all navigation. It established a real-time learned pipeline and clarified which parts of the problem were truly about deployment.

What worked

A learned visual controller could be integrated into a real-time Go1 stack.
The project exposed the importance of runtime smoothing, clipping, and conservative post-processing.
It made low-level movement and reactive safety a concrete engineering problem rather than only a training problem.

Limitations

The behavior remained mostly reactive.
Humans were not represented as explicit social agents.
The controller did not reason directly about crossing direction, yielding behavior, or ambiguity.

Why this led to the next study

The limitation of this phase was not only control. It was representation. The robot could learn motion patterns, but it still lacked a way to express social navigation decisions around people.

Earlier Robot Demo

This media belongs to the earlier real-time navigation study. The VLM work that followed is represented mainly by offline evaluation rather than a comparable online demo.

Go1 Walking Demo

Related Notes

The full archive now lives in the shared Research Notebook. The links here stay focused on the phase-specific notes most relevant to this study.