The coordinated dataflow and current-view target-conditioned execution as effective mechanisms for efficient persistent foundation-model navigation on power-constrained edge robots are established.
Abstract
Persistent open-vocabulary navigation enables robots to search for sequential language-specified objects using reusable visual semantic evidence. On edge GPUs, the pipeline is constrained by foundation-model inference, semantic projection, persistent map movement, frontier processing, and repeated target detection. Using OneMap as a controlled reference, we reorganize this workload through compiled mixed-precision perception, direct-to-map confidence-weighted feature aggregation, accelerator-resident semantic maps, compiled navigation operations, and language-conditioned detector scheduling. The resulting dataflow removes the dense image resolution feature intermediate and schedules target confirmation detection from current-view target similarity. On the complete Habitat-Matterport 3D benchmarks, our implementation retains 92.6% and 96.1% of the reproduced OneMap single- and multi-object success rates. Across 36 paired hardware-in-the-loop comparisons on a Jetson AGX Orin at four power modes, it achieves geometric mean speedups of 4.64× for on-device computation and 3.55× for the complete episode duration. It also reduces peak system random access memory (RAM) by 51.21% and a summed board-rail energy proxy by 75.32% on average while preserving every paired outcome. These results establish the coordinated dataflow and current-view target-conditioned execution as effective mechanisms for efficient persistent foundation-model navigation on power-constrained edge robots.
NPU offloading makes the Pi a much more efficient VLM platform: it cuts energy per generated token by about two-thirds, leaves the host CPU largely free, and avoids the thermal throttling that the CPU backend reaches in sustained use.
Jun-Lin Liao, Harshit Kondle, Di Liu et al.· 0 citations
Vision-language navigation (VLN) models perform well but target compute-rich platforms, limiting deployment on memory- and power-constrained robotic edge devices. Compression alone does not establish whether a VLN model fits the memory, latency, and energy budgets of an edge platform while preserving navigation behavio...
Rithvik Jonna, Man Namgung, Aakash Gurram et al.· arXiv.org· 0 citations
This work builds a GPU library that emulates SC matrix multiplication at scale, exposes stream lengths as first-class kernel arguments, and evaluates SC end-to-end on image classification, object detection and instance segmentation, class-conditional image generation, and visual world-model planning.
Hao-Ran Jin, Kang-Qi Zhang, Ji-Rong Yang et al.· arXiv.org· 0 citations
GPU robotics lacks the reusable numerical infrastructure of mature CPU stacks, instead relying on compiler frameworks that introduce overhead or repeatedly reimplementing numerical libraries. To address this, we introduce GLASS (GPU Linear Algebra Simple Subroutines), a header-only CUDA C++ library that provides thread...
GSPIM is presented, a temporal-locality-aware 4DGS accelerator for PIM-enabled edge devices that achieves a 4.15x geometric-mean speedup and a 4.99x geometric-mean energy-efficiency improvement over a fully optimized frame-by-frame GPU baseline.
Zi-Rui Ma, Zhi-Hua Fan, Hai-Bin Wu et al.· 0 citations
FieldRT is presented, an algorithm-hardware co-design methodology that uses a common analysis of geometric sensitivity and data lifetime to derive both the deployment profile and its spatial realization, and translates the retained geometric structure into a parameterized streaming architecture that shares field access...
Hai-Chuan Li· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.