Recoleta
GitHub
Source note

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

Authors
Kai Yang, Zedong Chu, Yingnan Guo, Zhengbo Wang, Shichao Xie, Yanfen Shen, et al.
Source
arXiv
Published
Collection
Embodied AI
Vision Language ActionMobile Robot NavigationCloud Edge ControlLatency CompensationSafe RL
Open arXivMarkdown source

Summary

AsyncShield is an edge-side adapter for cloud-based VLA robot navigation. It corrects delayed VLA waypoints with SE(2) pose realignment, then uses a safety-constrained RL policy to choose local sub-goals from LiDAR and corrected path inputs.

Problem

  • Cloud VLA models often run off-board, so network delay, jitter, packet loss, and outages make navigation commands arrive late.
  • A mobile robot keeps moving during that delay, so waypoints generated in an old ego frame can point to the wrong place in the current ego frame and cause collisions.
  • The problem matters because direct execution, temporal smoothing, and residual correction can fail when stale intents meet dynamic obstacles.

Approach

  • The edge device keeps a timestamped pose buffer and retrieves the robot pose at the VLA packet anchor time.
  • It maps each delayed VLA waypoint into the current ego frame with an analytic SE(2) transform, turning time lag into a measured spatial offset.
  • The adapter takes 5 realigned look-ahead waypoints spaced 0.2 m apart and 144 LiDAR proximity values.
  • A PPO-Lagrangian policy outputs a universal local sub-goal, with reward terms for path tracking and a LiDAR-based cost when the minimum obstacle distance falls below the safe radius.
  • Training uses randomized 10 m by 10 m scenes, 6 static and 6 dynamic obstacles, latency sampled from 0.3 to 1.5 s, packet loss up to 0.2, actuator lag, acceleration limits, noise, and angular bias.

Results

  • In 600 evaluation episodes under ideal network conditions, AsyncShield reached 80.0% success rate, 0.717 m CTE, 1.2% risk exposure rate, and 8.87 s time-to-goal. A2C2 reached 56.7% SR, RTC 40.0%, and Naive 20.0%.
  • Under mixed degradation, AsyncShield reached 76.7% SR, 0.725 m CTE, 1.3% RER, and 9.29 s TTG. A2C2 dropped to 43.3% SR, RTC to 30.0%, and Naive to 16.7%.
  • Under mixed degradation, AsyncShield reduced CTE versus A2C2 from 1.146 m to 0.725 m and versus Naive from 1.272 m to 0.725 m.
  • The ablation without temporal alignment fell from 76.7% to 36.7% SR under mixed degradation, with CTE rising from 0.725 m to 1.194 m.
  • The ablation without the RL adapter reached 53.3% SR and 1.443 m CTE under mixed degradation, which supports the need for learned intent-safety balancing.
  • The ablation without safety constraints reached only 23.3% SR under mixed degradation, with 4.7% RER despite low 0.692 m CTE, showing that close path tracking can be unsafe when commands are stale.

Link