Uncertainty-Aware Opportunistic Hybrid Language Model in Wireless Robotic Systems

In robotics, the promise of on-device language models hinges on their ability to generate reliable task plans without constant remote intervention. Yet real-world wireless environments introduce trade-offs between latency and accuracy. This study resolves this by validating an uncertainty-aware hybrid language model (U-HLM) on a fully instrumented robotic testbed, demonstrating its effectiveness in maintaining high-quality plan execution under fluctuating network conditions.

Rather than blindly invoking uplink transmission for remote LLM’s verification and resampling, U-HLM leverages uncertainty, the model’s self-assessed confidence in its outputs, to skip remote validation when the local model is confident. In trials with a multi-axis robotic arm, the robotic system was reliably provided with task plans, only consulting the remote server when draft reliability dipped. Qualitative observations showed that the accuracy of generated task plans dropped only to a negligible extent, while significantly reducing latency overhead from network round trips.

Not only that, U-HLM was not affected by the fluctuating network conditions, as much as the baseline methods did. While pure HLM, which does not leverage the uncertainty-aware opportunistic hybrid inference, suffered from significantly increased latency overhead under degraded network conditions, U-HLM was able to adapt as it is capable of selectively utilizing the wireless networks.

By grounding hybrid inference in real-world experimentation, this work provides a blueprint for uncertainty-aware robotics. Future research will explore scaling the validation to multi-robot fleets and integrating emerging 5G/6G edge infrastructures, further cementing U-HLM’s role in low-latency, high-reliability robotic systems.

[1] J. Park, Y. Lim, S. Oh, J. Park, J. Choi, and S. -L. Kim, “Uncertainty-Aware Opportunistic Hybrid Language Model in Wireless Robotic Systems,” in Machine Learning for Wireless Communication and Networks (ML4Wireless) Workshop, in Proc. of the 42nd International Conference on Machine Learning (ICML), Vancouver, BC, Canada, July 18, 2025.

[2] S. Oh*, J. Kim*, J. Park, S.-W. Ko, T. Q.S. Quek, and S.-L. Kim, "Uncertainty-Aware Hybrid Inference with On-Device Small and Remote Large Language Models," IEEE International Conference on Machine Learning for Communication and Networking (ICMLCN), Barcelona, Spain, July, 2025; Extended version submitted to IEEE JSAC.