BROWSE
Detail
Conference Paper
One-Step Planner: Unified Observation and Decision-Making with Vision-Language Models
Cited - time in 
Share 
Share
- Authors
- Youngjae Yoo, Jae-Woo Choi, Dohyung Kim, Byoung-Tak Zhang
- Issue Date
- 2026-06
- Citation
- International Conference on Robotics and Automation (ICRA) 2026 : Workshop, pp.1-4
- Publisher
- ICRA 2026 RIGOROUS Robot Perception Workshop
- Language
- English
- Type
- Conference Paper
- Abstract
- Most LLM-based embodied task planning systems adopt a modular pipeline that separates perception from planning: a vision model first produces a textual scene descrip- tion, and a language model then determines the next action. This separation introduces an information bottleneck in which rich visual input is compressed into text, discarding spatial relationships and fine-grained appearance cues. Moreover, the decoupled architecture introduces redundant cross-modal pro- cessing, where visual information is encoded into text and then reinterpreted for planning, leading to increased latency and memory overhead. We present One-Step Planner, a vision- language model (VLM) that unifies perception and planning in a single forward pass. By applying Low-Rank Adaptation (LoRA), instruction tuning creates a direct perception-to-action pathway that sharpens spatial attention toward task-relevant objects. We evaluate on two partially observable benchmarks, WAH-NL++ (a modified version of WAH-NL) and ALFRED, and systematically compare the end-to-end architecture against modular pipelines equipped with four different observer types. One-Step Planner achieves up to 10% point higher task success rates while reducing GPU memory usage by up to 44.6%. Spatial-attention further shows that the model focuses on goal- critical regions.
- KSP Keywords
- Decision-making, Decoupled architecture, Direct Perception, End to End(E2E), Fine grained(FG), GPU memory usage, Language Models, Low Rank, Memory overhead, Planning system(CPPS), Spatial attention