Saved in:
Bibliographic Details
Main Authors: Kresse, Fabian, Lampert, Christoph H.
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2511.07046
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866909907734757376
author Kresse, Fabian
Lampert, Christoph H.
author_facet Kresse, Fabian
Lampert, Christoph H.
contents Deploying continuous-control reinforcement learning policies on embedded hardware requires meeting tight latency and power budgets. Small FPGAs can deliver these, but only if costly floating point pipelines are avoided. We study quantization-aware training (QAT) of policies for integer inference and we present a learning-to-hardware pipeline that automatically selects low-bit policies and synthesizes them to an Artix-7 FPGA. Across five MuJoCo tasks, we obtain policy networks that are competitive with full precision (FP32) policies but require as few as 3 or even only 2 bits per weight, and per internal activation value, as long as input precision is chosen carefully. On the target hardware, the selected policies achieve inference latencies on the order of microseconds and consume microjoules per action, favorably comparing to a quantized reference. Last, we observe that the quantized policies exhibit increased input noise robustness compared to the floating-point baseline.
format Preprint
id arxiv_https___arxiv_org_abs_2511_07046
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Learning Quantized Continuous Controllers for Integer Hardware
Kresse, Fabian
Lampert, Christoph H.
Machine Learning
Artificial Intelligence
Deploying continuous-control reinforcement learning policies on embedded hardware requires meeting tight latency and power budgets. Small FPGAs can deliver these, but only if costly floating point pipelines are avoided. We study quantization-aware training (QAT) of policies for integer inference and we present a learning-to-hardware pipeline that automatically selects low-bit policies and synthesizes them to an Artix-7 FPGA. Across five MuJoCo tasks, we obtain policy networks that are competitive with full precision (FP32) policies but require as few as 3 or even only 2 bits per weight, and per internal activation value, as long as input precision is chosen carefully. On the target hardware, the selected policies achieve inference latencies on the order of microseconds and consume microjoules per action, favorably comparing to a quantized reference. Last, we observe that the quantized policies exhibit increased input noise robustness compared to the floating-point baseline.
title Learning Quantized Continuous Controllers for Integer Hardware
topic Machine Learning
Artificial Intelligence
url https://arxiv.org/abs/2511.07046