Skip to content

SVT-AV1 Reinforcement Learning Environment

UBC Capstone · TL26

Learning-based rate control through a direct connection between reinforcement learning and video encoding.

Code Documentation

The encoding control loop

Policy

Stable-Baselines3

Choose QP offsets

SVT-AV1 encoder

Per-superblock control

Encode video frames

Feedback

Observations and rewards

Video-quality metrics

Gymnasium environment · Python–C/C++ bridge
The environment connects policy actions to encoder callbacks and returns observations and feedback for learning.

Overview

A Gymnasium environment and control pipeline for the SVT-AV1 encoder, integrating Stable-Baselines3 for learning-based rate control and video quality optimization.

Method

Fine-grained rate control

The research fork exposes per-superblock QP offsets, allowing policies to influence encoding through callbacks.

Feedback across the language boundary

A Python–C/C++ bridge carries observations, actions, and feedback between the environment and encoder. PSNR and SSIM support video-quality evaluation.

A familiar training interface

Av1Env uses Gymnasium’s environment interface, with a Stable-Baselines3-compatible actor–critic policy, feature extractor, and training code.