[Fix] Fix issues noticed during Regression benchmark

README.md

@@ -49,9 +49,10 @@ The toolbox supports the most advanced high-quality navigation dataset, InternDa
 - [🏠 Introduction](#-introduction)
 - [🔥 News](#-news)
 - [📚 Getting Started](#-getting-started)
-- [📦 Overview of Benchmark \& Model Zoo](#-overview-of-benchmark-and-model-zoo)
+- [📦 Overview of Benchmark \& Model Zoo](#-overview)
 - [🔧 Customization](#-customization)
 - [👥 Contribute](#-contribute)
+- [🚀 Community Deployment & Best Practices](#-community-deployment--best-practices)
 - [🔗 Citation](#-citation)
 - [📄 License](#-license)
 - [👏 Acknowledgements](#-acknowledgements)
@@ -213,6 +214,23 @@ For example, raising issues, fixing bugs in the framework, and adapting or addin
 
 **Note:** We welcome the feedback of the model's zero-shot performance when deploying in your own environment. Please show us your results and offer us your future demands regarding the model's capability. We will select the most valuable ones and collaborate with users together to solve them in the next few months :)
 
+## 🚀 Community Deployment & Best Practices
+
+We are excited to see InternNav being deployed and extended by the community across different robots and real-world scenarios.  
+Below are selected community-driven deployment guides and solution write-ups, which may serve as practical references for advanced users.
+
+- **IROS Challenge Nav Track: Champion Solution (2025)**  
+  A complete system-level solution and design analysis for Vision-and-Language Navigation in Physical Environments.  
+  🔗 https://zhuanlan.zhihu.com/p/1969046543286907790
+
+- **Go2 Series Deployment Tutorial (ShanghaiTech University)**  
+  Step-by-step edge deployment guide for InternNav-based perception and navigation.  
+  🔗 https://github.com/cmjang/InternNav-deploy
+
+- **G1 Series Deployment Tutorial (Wuhan University)**  
+  Detailed educational materials on vision-language navigation deployment.  
+  🔗 [*Chapter 5: Vision-Language Navigation (Part II)*](https://mp.weixin.qq.com/s/p3cJzbRvecMajiTh9mXoAw)
+
 ## 🔗 Citation
 
 If you find our work helpful, please cite:

internnav/agent/dialog_agent.py

@@ -11,21 +11,16 @@
 import quaternion
 import torch
 from PIL import Image, ImageDraw
-from transformers import (
-    AutoProcessor,
-    AutoTokenizer,
-    Qwen2_5_VLForConditionalGeneration,
-)
 
 from internnav.agent import Agent
 from internnav.configs.agent import AgentCfg
 
 try:
-    pass
-except Exception as e:
-    print(f"Warning: ({e}), Ignore this if not using dual_system.")
-
-try:
+    from transformers import (
+        AutoProcessor,
+        AutoTokenizer,
+        Qwen2_5_VLForConditionalGeneration,
+    )
     from depth_camera_filtering import filter_depth
     from habitat.tasks.nav.shortest_path_follower import ShortestPathFollower
 except Exception as e:

internnav/dataset/vlln_lerobot_dataset.py

@@ -22,21 +22,21 @@
 
 # Define placeholders for dataset paths
 IION_split1 = {
-    "data_path": "traj_data/mp3d_split1",
+    "data_path": "projects/VL-LN-Bench/traj_data/mp3d_split1",
     "height": 125,
     "pitch_1": 0,
     "pitch_2": 30,
 }
 
 IION_split2 = {
-    "data_path": "traj_data/mp3d_split2",
+    "data_path": "projects/VL-LN-Bench/traj_data/mp3d_split2",
     "height": 125,
     "pitch_1": 0,
     "pitch_2": 30,
 }
 
 IION_split3 = {
-    "data_path": "traj_data/mp3d_split3",
+    "data_path": "projects/VL-LN-Bench/traj_data/mp3d_split3",
     "height": 125,
     "pitch_1": 0,
     "pitch_2": 30,

internnav/evaluator/utils/result_logger.py

@@ -319,5 +319,5 @@ def finalize_all_results(self, rank, world_size):
             }
 
         # write log content to file
-        with open(f"{self.name}_result.json", "w") as f:
+        with open(f"{PROJECT_ROOT_PATH}/logs/{self.name}/result.json", "w") as f:
             json.dump(json_data, f, indent=2, ensure_ascii=False)

internnav/habitat_extensions/vlln/README.md

@@ -3,16 +3,34 @@
 Vision-Language-and-Language Navigation (VL-LN) is a new [benchmark](https://0309hws.github.io/VL-LN.github.io/) built upon VLN in Habitat, which refers to the setting that models take the vision and language as input and output language and navigation actions. In contrast to VLN, where agents only take navigation actions, agents in VL-LN could ask questions and engage in dialogue with users to complete tasks better with language interaction. 
 This package adapts [Meta AI Habitat](https://aihabitat.org) for VL-LN within InternNav. It wraps Habitat environments that expose semantic masks, registers dialog-aware datasets and metrics, and provides evaluators that coordinate agent actions, NPC interactions, and logging.
 
+Install our benchmark [dataset](https://huggingface.co/datasets/InternRobotics/VL-LN-Bench) and the [latest checkpoints](https://huggingface.co/InternRobotics/VL-LN-Bench-basemodel) from HuggingFace.
+Place the downloaded benchmark under `InternNav/projects/VL-LN-Bench` to match the default path expected by the code.
+
 ## Package structure
 
 ```
-habitat_vlln_extensions/
-├── __init__.py
-├── habitat_dialog_evaluator.py
-├── habitat_vlln_env.py
-├── measures.py
-├── simple_npc/
-└── utils/
+InternNav
+├── assets/
+├── internnav/
+│   ├── habitat_vlln_extensions
+│   │   ├── simple_npc
+│   │   │   ├── api_key.txt
+│   │   ├── measures.py
+│   │   ├── habitat_dialog_evaluator.py
+│   │   ├── habitat_vlln_env.py
+│   ... ... ...
+...
+├── projects
+│   ├── VL-LN-Bench/
+│   │   ├── base_model/ 
+│   │   ├── raw_data/ 
+│   │   ├── scene_datasets/
+│   │   │   └── mp3d/
+│   │   │       └── 17DRP5sb8fy/
+│   │   │       ├── 1LXtFkjw3qL/
+│   │   │       ...
+│   │   ├── traj_data/
+...
 ```
 
 * `__init__.py` re-exports the public entry points so callers can import

internnav/habitat_extensions/vlln/habitat_dialog_evaluator.py

@@ -257,6 +257,13 @@ def calc_metrics(self, global_metrics: dict) -> dict:
         # avoid /0 if no episodes
         denom = max(len(sucs_all), 1)
 
+        # clean NaN in spls, treat as 0.0
+        torch.nan_to_num(spls_all, nan=0.0, posinf=0.0, neginf=0.0, out=spls_all)
+
+        # clean inf in nes, only fiinite nes are counted
+        nes_finite_mask = torch.isfinite(nes_all)
+        nes_all = nes_all[nes_finite_mask]
+
         return {
             "sucs_all": float(sucs_all.mean().item()) if denom > 0 else 0.0,
             "spls_all": float(spls_all.mean().item()) if denom > 0 else 0.0,

internnav/habitat_extensions/vln/README.md

@@ -9,7 +9,7 @@ utilities.
 ## Package structure
 
 ```
-habitat_extensions/
+habitat_extensions/vln/
 ├── __init__.py
 ├── habitat_env.py
 ├── habitat_default_evaluator.py

internnav/habitat_extensions/vln/habitat_vln_evaluator.py

@@ -192,7 +192,7 @@ def eval_action(self):
             "nes": nes,  # shape [N_local]
         }
 
-        if ndtws:
+        if ndtws is not None:
             result["ndtws"] = ndtws  # shape [N_local]
         return result
 
@@ -207,6 +207,13 @@ def calc_metrics(self, global_metrics: dict) -> dict:
 
         # avoid /0 if no episodes
         denom = max(len(sucs_all), 1)
+
+        # clean NaN in spls, treat as 0.0
+        torch.nan_to_num(spls_all, nan=0.0, posinf=0.0, neginf=0.0, out=spls_all)
+
+        # clean inf in nes, only fiinite nes are counted
+        nes_finite_mask = torch.isfinite(nes_all)
+        nes_all = nes_all[nes_finite_mask]
 
         result_all = {
             "sucs_all": float(sucs_all.mean().item()) if denom > 0 else 0.0,
@@ -587,7 +594,7 @@ def _run_eval_dual_system(self) -> tuple:
             torch.tensor(spls).to(self.device),
             torch.tensor(oss).to(self.device),
             torch.tensor(nes).to(self.device),
-            torch.tensor(ndtw).to(self.device) if 'ndtw' in metrics else None,
+            torch.tensor(ndtw).to(self.device) if ndtw else None,
         )
 
     def _run_eval_system2(self) -> tuple:
@@ -876,5 +883,5 @@ def _run_eval_system2(self) -> tuple:
             torch.tensor(spls).to(self.device),
             torch.tensor(oss).to(self.device),
             torch.tensor(nes).to(self.device),
-            torch.tensor(ndtw).to(self.device) if 'ndtw' in metrics else None,
+            torch.tensor(ndtw).to(self.device) if ndtw else None,
         )