arXiv:2611.18035v1 Announce Type: cross
Abstract: On-device reinforcement learning runs a periodic actor and a learner on one embedded GPU shared with other workloads. Because the learner produces the actor's weights, existing designs stall the actor or let its policy drift. We present RL$^{\mathrm{RT}}$, an on-device runtime that makes this coupling a tunable parameter: a refresh bound limits actor releases between refreshes, and a GPU permit with slack-aware admission prioritises the actor and fills spare time with training atop any base scheduler. Under the real refresh cost on two Jetson boards, RL$^{\mathrm{RT}}$ keeps a 16.8 M-parameter policy's deadline misses below 1%, against 85 to 100% without it.
arxiv
RL^RT: On-Device Reinforcement Learning Runtime
Researchers present RL^RT, an on-device runtime for reinforcement learning in which a periodic actor and a learner share one embedded GPU with other workloads. The runtime makes actor-learner coupling tunable through a refresh bound that limits actor releases between refreshes and a slack-aware GPU permit that prioritizes the actor while filling spare time with training on top of any base scheduler. Evaluated under real refresh cost on two Jetson boards, it keeps a 16.8M-parameter policy's deadline misses below 1%, compared with 85% to 100% without it.
This paper offers a concrete scheduling approach and measured embedded-GPU results for running RL actor and learner together on device, useful to practitioners building on-device RL systems.
Comments & Questions (0)
No comments yet
Be the first to comment!