Three Amap Papers Accepted at IROS 2026: VLN Navigation, VLA Latency Correction, and Diffusion‑Based Quadruped Control

Explore Like Humans: Autonomous Exploration with Online SG‑Memo Construction for Embodied Agents

Embodied agents navigating complex indoor spaces need structured spatial memory for reasoning and decision‑making. Existing methods separate exploration and offline memory reconstruction, relying on geometry alone and often missing semantic landmarks such as doorways and stairs, which reduces exploration efficiency and environment coverage.

The ABot‑Explorer framework unifies memory construction and active exploration into a single online stage that uses only RGB input. It extracts Semantic Navigational Affordances (SNA) from observations via a large visual‑language model, treating SNA as human‑aligned anchors that prioritize structural transition nodes (e.g., corridor entrances, stairways). SNA is dynamically integrated into a hierarchical SG‑Memo (scene‑graph memory), enabling simultaneous exploration and mapping.

Key contributions are: (1) the first online, unified active‑exploration‑and‑memory‑construction framework; (2) a large‑scale SNA and SG‑Memo annotation dataset built with InteriorGS; and (3) experiments showing ABot‑Explorer significantly outperforms state‑of‑the‑art methods in exploration efficiency and environment coverage, while the generated SG‑Memo supports multiple downstream tasks.

AsyncShield: A Plug‑and‑Play Edge Adapter for Asynchronous Cloud‑based VLA Navigation

Large visual‑language‑action (VLA) models provide strong zero‑shot generalization for robot navigation but must run in the cloud due to their size. Cloud deployment inevitably introduces network jitter and inference latency, causing the returned action commands to correspond to past observations and leading to spatiotemporal misalignment, collisions, and task failures. Existing approaches rely on black‑box time‑series prediction, which lacks physical interpretability and generalization.

AsyncShield is a lightweight, plug‑and‑play edge adapter that corrects stale cloud commands through deterministic spatial mapping rather than temporal prediction. It uses a time‑pose buffer and kinematic transforms to convert communication delay into a computable spatial offset (a white‑box physical mapping). The edge adapter is modeled as a Constrained Markov Decision Process (CMDP) and trained with PPO‑Lagrangian reinforcement learning to balance intent fidelity and safety, while high‑frequency LiDAR hard constraints ensure physical safety.

Design elements such as standardized sub‑goal interfaces, domain randomization, and collision‑radius inflation allow AsyncShield to achieve zero‑shot transfer without fine‑tuning any cloud‑based foundation model. Simulation and real‑world experiments demonstrate a substantial increase in asynchronous navigation success rate, physical safety, and robust zero‑shot generalization.

Constraint‑Aware Diffusion Priors for High‑Fidelity and Versatile Quadruped Locomotion

Combining reinforcement learning and imitation learning has advanced bio‑inspired quadruped control, yet scaling to large, heterogeneous datasets reveals fundamental bottlenecks: GAN‑based discriminators suffer mode collapse and cannot model diverse motion distributions; traditional kinematic priors produce tracking conflicts under out‑of‑distribution conditions, causing unintended drift; and unconstrained priors ignore actuator dynamics, creating safety risks on hardware.

Diff‑CAST (Diffusion‑guided Constraint‑Aware Symmetric Tracking) addresses these issues with a diffusion‑model‑based motion prior that generates stylized rewards, replacing GAN discriminators and enabling robust data expansion across heterogeneous datasets. To ensure high‑fidelity intent execution and safe real‑world deployment, a complete Sim2Real pipeline is built, featuring a Symmetric Augmented Command Conditioning (SACC) module for drift‑free tracking and a Constrained Reinforcement Learning module that respects hardware limits.

Quadruped experiments show Diff‑CAST mitigates mode collapse, achieves smooth transitions among multiple skills, and delivers robust motions that satisfy hardware constraints, confirming its effectiveness for high‑fidelity, versatile locomotion.