FoveaRL · foveated streaming for VR robot teleoperation

A reinforcement-learning agent decides, 5 times a second, how many bits the operator's eyes actually need.

● real trained DQN agents held-out seeds 10000-10003 (not hand-picked) simulation replay · not hardware

Network

Run

0.0 s

Arena · same course, same network, six controllers

Controller	Progress	Collisions	Frame age	Quality Q	Sent	Status

LIVE FEED

frame age

fovea

quality Q

0.00

Fovea radius r

Periphery res. p

Bitrate b

Network latency120 ms

Move the mouse over the view: it is your gaze. The sharp circle follows your gaze one latency late — exactly the mismatch the agent has to manage.

Live telemetry

Replays of the paper's simulator (quality is a parametric proxy, operator & network models are assumed). Built with Three.js · all data embedded, works offline.