Eku127/swiftvln-satnav-3b-1ep-f32s4-overlap0-pf-h8-pool-s2-noembed
Eku127/swiftvln-satnav-3b-1ep-f32s4-overlap0-pf-h8-pool-s2-noembed is a 3 billion parameter vision-and-language navigation model, fine-tuned from Qwen2.5-VL-3B-Instruct for the SatNav benchmark. It predicts short navigation-action sequences using an instruction, current RGB window, and sampled visual memory, specifically designed for continuous-state navigation over satellite imagery. This model is optimized for tasks requiring visual navigation and sequential decision-making in complex environments, leveraging a 32-frame context window and per-frame memory.
Loading preview...
Overview
This model, Eku127/swiftvln-satnav-3b-1ep-f32s4-overlap0-pf-h8-pool-s2-noembed, is a specialized vision-and-language navigation (VLN) checkpoint developed by Eku127. It is fine-tuned from the Qwen2.5-VL-3B-Instruct backbone, specifically for the SatNav benchmark, which involves continuous-state navigation over satellite imagery. The model predicts short sequences of navigation actions based on an instruction, the current RGB visual input, and sampled visual memory.
Key Capabilities
- Vision-and-Language Navigation: Interprets natural language instructions to navigate through satellite imagery.
- Action Sequence Prediction: Generates short sequences of navigation actions (4 actions per prediction horizon).
- Visual Memory Integration: Utilizes a 32-frame RGB window and up to 8 uniformly sampled history frames for contextual awareness, employing per-frame average pooling with stride 2.
- Specialized Training: Fine-tuned for one epoch on SatNav-v0.1 offline expert trajectories, ensuring high relevance to satellite navigation tasks.
Good For
- Research in VLN: Ideal for researchers working on vision-and-language navigation, especially in satellite imagery domains.
- SatNav Benchmark Evaluation: Serves as a reference model for the SwiftVLN SatNav Ablation Model Zoo and can be used with the SwiftVLN evaluation guide.
- Developing Navigation Agents: Suitable for building and testing agents that require interpreting instructions and navigating complex visual environments.