Source-linked AI summary
ReMem-VLA: Empowering Vision-Language-Action Model with Memory via Dual-Level Recurrent Queries
Hang Li, Fengyi Shen, Dong Chen, Liudi Yang, Xudong Wang, Jinkui Shi, Zhenshan Bing, Ziyuan Liu, Alois Knoll
TL;DR
ReMem-VLA addresses the limited historical context of Markovian VLA control with dual-level recurrent memory queries and Past Observation Prediction. Across simulation and real-robot memory-demanding tasks, it demonstrates strong capabilities spanning spatial, temporal, visual, episodic, and sequential memory.
Problem
Markovian VLA control neglects temporal causality, causing failures on manipulation tasks that require remembering prior states and actions.
Method
ReMem-VLA uses frame-level and chunk-level recurrent memory queries for short- and long-term retention, augmented by Past Observation Prediction for visual memory.
Results
ReMem-VLA demonstrates strong memory capability across spatial, temporal, visual, episodic, and sequential manipulation tasks in simulation and real-robot settings.
Takeaways & Limitations
Dual-level recurrence supports multi-timescale memory, while Past Observation Prediction particularly strengthens performance on visually memory-intensive tasks.
Takeaways & Limitations
The model has not been trained on large-scale robot datasets, which may affect its generalization ability.
Abstract
from arXiv · showhide
Vision-language-action (VLA) models for closed-loop robot control are typically cast under the Markov assumption, making them prone to errors on tasks requiring historical context. To incorporate memory, existing VLAs either retrieve from a memory bank, which can be misled by distractors, or extend the frame window, whose fixed horizon still limits long-term retention. In this paper, we introduce ReMem-VLA, a Recurrent Memory VLA model equipped with two sets of learnable queries: frame-level recurrent memory queries for propagating information across consecutive frames to support short-term memory, and chunk-level recurrent memory queries for carrying context across temporal chunks for long-term memory. These queries are trained end-to-end to aggregate and maintain relevant context over time, implicitly guiding the model's decisions without additional training or inference cost. Furthermore, to enhance visual memory, we introduce Past Observation Prediction as an auxiliary training objective. Through extensive memory-centric simulation and real-world robot experiments, we demonstrate that ReMem-VLA exhibits strong memory capabilities across multiple dimensions, including spatial, sequential, episodic, temporal, and visual memory. ReMem-VLA significantly outperforms memory-free VLA baselines $π$0.5 and OpenVLA-OFT and surpasses MemoryVLA on memory-dependent tasks by a large margin.
1 Introduction
ReMem-VLA addresses the failure of Markovian VLA control on memory-dependent manipulation by combining recurrent memory at frame and chunk timescales with past observation prediction. It is trained and evaluated to retain short- and long-horizon task context across multiple memory dimensions.
- Motivation: Markovian VLAs neglect temporal causality, causing failures when manipulation requires recalling prior states or environmental changes.The paper targets memory within single task episodes, spanning short- and long-horizon tasks rather than lifelong memory.
- Approach: ReMem-VLA uses frame-level recurrent queries for short-term memory and chunk-level recurrent queries for longer-term retention.Frame-level queries propagate context across frames, while chunk-level queries accumulate information over longer intervals.
- Training: Slot-based streaming training batches variable-length episodes while preserving temporal continuity and preventing cross-episode state mixing.Each slot tracks one live episode and contributes one frame per training step.
- Evaluation: ReMem-VLA shows consistent and substantial gains over baselines across spatial, temporal, sequential, episodic, and visual memory tasks in simulation and real-world experiments.Ablations identify dual-level recurrence as critical for memory retention and confirm the contribution of Past Observation Prediction.
- Approach: Past Observation Prediction strengthens visual memory by training the model to recover historical visual information.The model uses hindsight queries for past image reconstruction alongside action prediction.
2 Related Work
Prior visuomotor policies incorporate history through longer observation windows, sparse representations, or memory-bank retrieval. These strategies remain constrained by fixed horizons, omitted details, external-model dependence, or retrieval interference.
- VLA approaches: Continuous-action VLAs add regression, diffusion, or flow-matching heads to predict continuous action trajectories.The related systems include π0, π0.5, OpenVLA-OFT, GR00T-N1, and InternVLA-M1.
- Historical memory: Extended-window methods condition action prediction on multiple historical frames but remain limited when tasks exceed their fixed horizon.Examples include CronusVLA, HAMLET, PTP, and PAM.
- Historical memory: Sparse history representations encode past motion or selected keyframes but can omit essential details and depend on external model stability.TraceVLA, HistRISE, and Bpp represent different sparse-history strategies.
- Historical memory: Memory-bank retrieval can process historical information but depends heavily on strong cues and suffers interference at scale.The paper contrasts this limitation with its goal of a unified memory mechanism.
3 Method
ReMem-VLA extends a history-aware VLA with dual-level recurrent memory queries that retain context at frame and chunk timescales. A connector fuses current and historical representations for action-chunk generation and past-image reconstruction, while fixed EMA recurrence supports long-horizon propagation.
- Problem formulation: ReMem-VLA augments Markovian action prediction with recurrent memory state and auxiliary past-observation prediction.The history-aware formulation predicts both future action chunks and a reconstructed past observation from current inputs, language, and recurrent memory.
- Dual-level recurrent memory: Dual-level queries provide frame-level short-term retention and chunk-level long-term retention through separate recurrent update paths.Frame-level memory updates every timestep, whereas chunk-level memory updates at chunk boundaries.
- Architecture and prediction: A bidirectional-attention connector lets action and hindsight queries interact with recurrent memory queries before prediction heads generate actions and reconstruct past images.The architecture uses a frozen VLM backbone, task-specific action and hindsight queries, and prediction heads for action generation and past-image reconstruction.
- Frame-level memory: Frame-level memory uses dense EMA propagation to track recent object states, ongoing actions, and immediate temporal dependencies.The fixed coefficient βf controls the frame-level exponential moving average update.
- Chunk-level memory: Chunk-level memory updates every k frames, producing slower decay than frame-level memory for stable retention of initial configurations and long temporal dependencies.Its decay is described as (1−βc)n/k compared with frame-level (1−βf)n.
- Training recurrence: The gradient-free recurrence combines a frozen VLM with fixed EMA fusion, avoiding long-trajectory backpropagation while preserving deterministic memory propagation.Streaming slot-based batching maintains recurrent state across each episode and resets it at episode boundaries to prevent cross-episode leakage.
4 Experiments
ReMem-VLA is evaluated on extended simulation and real-world memory-dependent manipulation tasks, with ablations testing recurrent-query levels, visual-memory prediction, recurrent updates, EMA settings, and query counts.
- Simulation Experiments: MemoryBench includes three approximately 300-frame tasks and a fourth long-horizon task exceeding 600 frames.The added task combines Put Block Back and Rearrange Block to test longer memory retention.
- Simulation Experiments: ReMem-VLA achieves the best performance across all extended MemoryBench tasks compared with baselines.The evaluation compares against OpenVLA-OFT, π0.5, and MemoryVLA under a shared protocol.
- Real World Experiments: ReMem-VLA achieves an 82.5% average real-world success rate, outperforming MemoryVLA by 8% and π0.5 by 11%.The real-world tasks probe visual, spatial, temporal, and episodic memory.
- Ablation Study: Removing recurrent queries collapses memory capability, while dual-level queries outperform frame-only and chunk-only variants across MemoryBench tasks.Frame-level queries support short-term memory, whereas chunk-level queries improve long-term retention; the combined model supplies complementary capabilities.
- Ablation Study: Past Observation Prediction improves Return Fruit success from 34% to 82%, while offering limited benefit on tasks without visual-memory requirements.The result indicates that recurrent queries alone preserve visual details less effectively than other memory dimensions.
- Ablation Study: A moderate query count of 128 performs best, β = 0.5 gives the best EMA performance, and a 1× action-chunk update interval yields the best overall result.Trainable recurrent paths nearly eliminate memory capability, while excessive queries can add redundancy and attention noise.
5 Conclusion
ReMem-VLA combines frame-level and chunk-level recurrence to provide short-term and long-horizon memory, with Past Observation Prediction enhancing visual memory. It demonstrates memory capabilities across multiple dimensions in simulated and real-robot manipulation tasks, while large-scale robot-data training remains an open limitation.
- Frame-level recurrence provides short-term memory, while chunk-level recurrence supports long-horizon retention.
- Past Observation Prediction improves ReMem-VLA’s visual memory.
- ReMem-VLA demonstrates spatial, temporal, visual, episodic, and sequential memory capabilities across simulation and real-robot tasks.
- Training on large-scale robot datasets remains an unresolved limitation that may affect generalization ability.