HeLM: Hierarchical Explicit Language Memory for Non-Markovian Manipulation
Abstract
A hierarchical VLA framework is proposed to address the limitations of Markovian models by maintaining task history through explicit language memory and event-triggered updates.
Recent advances in Vision-Language-Action (VLA) models have enabled fine manipulation capabilities, yet performing tasks that require working memory remains a key challenge. Most existing VLA architectures operate under a Markovian assumption, mapping only the current observation to an action, which leads to failure in non-Markovian environments. Inspired by the human 'inner monologue', we propose Hierarchical Explicit Language Memory (HeLM), a hierarchical VLA framework that maintains task-relevant history in explicit language through factorized event detection and event-triggered memory updates. We evaluate HeLM across 10 non-Markovian real-world tasks with 56 variants, including five newly formulated inter-episode tasks that require persistent context across task boundaries. HeLM improves intra-episode success by 12.8 percentage points over the strongest baseline and achieves 78% subtask success on compositional generalization tasks, while using up to 4.5x less memory and running 2.5x faster than the keyframe baseline at long horizons.
Community
🤖 HeLM: Hierarchical Explicit Language Memory for Non-Markovian Manipulation
How can robots track task progress when the current observation alone is insufficient?
We introduce HeLM, a hierarchical VLM–VLA framework that maintains task-relevant history in explicit language memory.
Two key aspects of our work:
🧠Event-triggered memory updates: HeLM separates visual event detection from memory updates. The HLP updates memory and generates the next action command only when a task-relevant event is detected or the instruction changes.
🔄 Inter-episode memory: We introduce tasks where robots must remember information acquired under previous instructions, such as a human-demonstrated button sequence.
We evaluate HeLM on 10 real-world non-Markovian tasks (56 variants). Our two-stage inference improves average intra-episode subtask success from 67.7% to 85.7% compared with a one-stage variant.
We also provide detailed failure analyses and evaluations of memory scalability.
We welcome feedback and discussions on memory for robot learning!
Models citing this paper 0
No model linking this paper
Datasets citing this paper 0
No dataset linking this paper
Spaces citing this paper 0
No Space linking this paper
Collections including this paper 0
No Collection including this paper