Research map / Vision-language and multimodal / Robot action and navigation

ECHO: Continuous Hierarchical Memory for Vision-Language-Action Models

NeurIPS 2026 · Vision-language and multimodal › Robot action and navigation · approach: Vision-language-action foundation models · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes ECHO, a hyperbolic (Lorentz) hierarchical memory module with entailment-cone tree retrieval and background consolidation, plugged into the pi0 VLA to improve long-horizon manipulation on LIBERO.

Problem

Existing memory-augmented VLA models store experiences as linear or flat banks, lacking hierarchical structure for manipulation categories, which hinders efficient retrieval and generalization to unseen long-horizon task compositions.

Approach

Maps VLA hidden states via a hyperbolic autoencoder into a Lorentz-manifold memory tree with entailment-cone constraints, enabling top-down beam-search retrieval, node splitting via Lorentzian K-Means, and geodesic interpolation to synthesize virtual memories, injected into pi0 through a similarity-modulated residual.

Key results

• On LIBERO-Long, ECHO raises success from 80.7% (Vanilla pi0) to 93.5%, a 12.8-point absolute gain.
• Cross-suite generalization on LIBERO-Long using only source-suite memories improves average success from 80.70% to 89.31%.
• Real-world average success rises from 58.3% to 70.0% across three Franka tasks (20 trials each).

Surprising findings

• Flat Euclidean memory yields only a modest gain (80.70% to 83.25%) with higher variance, while short-term buffer alone gives 88.81%.

Limitations

• Performance depends on quality of VLM-based subgoal extraction and relevance of retrieved memories.
• Real-world evaluation limited to three tabletop tasks with 20 trials each; described as preliminary.
• Several baseline results (Octo, OpenVLA, MAP-VLA) are taken from reported settings rather than controlled reproduction.

Experimental setup

Datasets: LIBERO LIBERO-Plus curated robotic task executions and semantic transition frames

Benchmarks: LIBERO-Spatial LIBERO-Object LIBERO-Goal LIBERO-Long LIBERO-Plus Real-world Franka tasks (Place Banana in Bowl, Stack Blocks, Insert Circle into Base)

Metrics: success rate (%) retrieval latency

Baselines: Octo OpenVLA MAP-VLA MemoryVLA Vanilla pi0 Flat memory (Euclidean kNN)

Builds on: pi0 Lorentz model of hyperbolic space Hyperbolic entailment cones Lorentzian K-Means

Closest papers by meaning

• EfficientVLA: Training-Free Acceleration and Compression for Vision-Language-Action Models (NeurIPS 2025 · similarity 0.89)
• Not All Features Are Created Equal: A Mechanistic Study of Vision-Language-Action Models (NeurIPS 2026 · similarity 0.88)
• VLA-Pro: Cross-Task Procedural Memory Transfer for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.88)
• SAM2Act: Integrating Visual Foundation Model with A Memory Architecture for Robotic Manipulation (ICML 2025 · similarity 0.88)
• VLA-OS: Structuring and Dissecting Planning Representations and Paradigms in Vision-Language-Action Models (NeurIPS 2025 · similarity 0.88)
• VisMem: Latent Vision Memory Unlocks Potential of Vision-Language Models (CVPR 2026 · similarity 0.88)