∞-THOR
收藏资源简介:
∞-THOR是一个用于长时序具身任务的新框架,旨在推动具身AI的长时序理解。该框架提供了合成可扩展、可重复和无限长时序轨迹的生成框架,以及一个新的具身QA任务,需要在扩展的轨迹中测试代理的长时序推理能力。此外,∞-THOR还提供了一个长时序数据集和基准测试套件,包括跨越数百个环境步骤的复杂任务,每个任务都配对有地面真实动作序列。这些数据集和代码可以在pearls-lab.github.io/infini-thor找到。
∞-THOR is a novel framework for long-horizon embodied tasks, aimed at advancing long-horizon comprehension of embodied AI. This framework provides a generative pipeline for synthesizing scalable, reproducible, and infinitely long-horizon trajectories, alongside a new embodied QA task designed to evaluate agents' long-horizon reasoning capabilities across extended trajectories. Furthermore, ∞-THOR also releases long-horizon datasets and benchmark suites, which contain complex tasks spanning hundreds of environmental steps, with each task paired with ground-truth action sequences. These datasets and code are available at pearls-lab.github.io/infini-thor.

- 1Beyond Needle(s) in the Embodied Haystack: Environment, Architecture, and Training Considerations for Long Context Reasoning加州大学圣地亚哥分校 · 2025年



