A GPU-Parallel Framework for Heterogeneous Multi-Task Reinforcement Learning

Name: A GPU-Parallel Framework for Heterogeneous Multi-Task Reinforcement Learning
Uploaded: 2026-10-06T05:07:43.342819+00:00
Duration: 1069 s
Description: The work introduces MT-Libero, a GPU-parallel framework that converts structured manipulation task families into multi-task reinforcement learning benchmarks by separating task semantics from execution. It proposes DGPO, an on-policy method combining importance weighted PPO and adaptive behavior cloning, to allow simultaneous learning across different tasks using expert demonstrations.

The work introduces MT-Libero, a GPU-parallel framework that converts structured manipulation task families into multi-task reinforcement learning benchmarks by separating task semantics from execution. It proposes DGPO, an on-policy method combining importance weighted PPO and adaptive behavior cloning, to allow simultaneous learning across different tasks using expert demonstrations.

Open the full episode