Upload folder using huggingface_hub

README.md

@@ -0,0 +1,88 @@
+---
+tags:
+  - robotics
+  - anima
+  - freya
+  - gaussian-splatting
+  - slam
+  - lidar
+  - 3d-reconstruction
+  - sensor-fusion
+  - robot-flow-labs
+library_name: pytorch
+pipeline_tag: robotics
+license: apache-2.0
+---
+
+# FREYA — Gaussian-LIC2 SLAM (ANIMA Module)
+
+Part of the [ANIMA Perception Suite](https://github.com/RobotFlow-Labs) by Robot Flow Labs.
+
+## Paper
+
+**Gaussian-LIC2: LiDAR-Inertial-Camera Gaussian Splatting SLAM** (arXiv:2507.04004)
+
+Xiaolei Lang, Jiajun Lv, Kai Tang, Laijian Li, Jianxin Huang, Lina Liu, Yong Liu, Xingxing Zuo
+
+## Architecture
+
+Tightly-coupled LiDAR+Inertial+Camera Gaussian Splatting SLAM:
+- Continuous-time B-spline SE(3) trajectory optimization
+- Incremental 3D Gaussian map with SH lighting (degree 3)
+- Zero-shot depth completion for LiDAR-blind areas
+- CUDA-accelerated tile-based rendering
+- Per-scene optimization (not batch training)
+
+## Exported Formats
+
+| Format | File | Use Case |
+|--------|------|----------|
+| PyTorch (.pth) | `pytorch/freya_v1.pth` | Training, fine-tuning |
+| SafeTensors | `pytorch/freya_v1.safetensors` | Fast loading, safe |
+| ONNX | `onnx/freya_depth_v1.onnx` | Cross-platform depth inference |
+| TensorRT FP16 | `tensorrt/freya_depth_v1_fp16.engine` | Edge deployment (Jetson/L4) |
+| TensorRT FP32 | `tensorrt/freya_depth_v1_fp32.engine` | Full precision inference |
+| Checkpoint | `checkpoints/best.pth` | Resume SLAM optimization |
+| Config | `configs/training.toml` | Reproducibility |
+| Logs | `logs/training_history.json` | Loss curves, metrics |
+
+## Training Details
+
+| Parameter | Value |
+|-----------|-------|
+| Hardware | NVIDIA L4 (23GB VRAM) |
+| Dataset | TUM VI Benchmark (room1, 2821 frames) |
+| Gaussians | 12.2M pre-allocated, 500K active |
+| Optimizer | Adam (per-param, foreach=False) |
+| Mixed Precision | bf16 |
+| Best Loss | 0.0010 |
+| VRAM Usage | 17.8GB (77%) |
+
+## Usage
+
+```python
+import torch
+
+state = torch.load("pytorch/freya_v1.pth", weights_only=False)
+gaussian_map = state["gaussian_map"]
+trajectory = state["trajectory"]
+
+# SafeTensors
+from safetensors.torch import load_file
+tensors = load_file("pytorch/freya_v1.safetensors")
+```
+
+## Citation
+
+```bibtex
+@article{lang2025gaussianlic2,
+  title={Gaussian-LIC2: LiDAR-Inertial-Camera Gaussian Splatting SLAM},
+  author={Lang, Xiaolei and Lv, Jiajun and Tang, Kai and Li, Laijian and Huang, Jianxin and Liu, Lina and Liu, Yong and Zuo, Xingxing},
+  year={2025},
+  journal={arXiv preprint arXiv:2507.04004}
+}
+```
+
+## License
+
+Apache 2.0 — Robot Flow Labs / AIFLOW LABS LIMITED

checkpoints/best.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:76b369ff2a2d531a71fea3c3672d0b46014e7e79dbb4b23fe4870b97cc6685e2
+size 147247

configs/training.toml

@@ -0,0 +1,53 @@
+# TUM VI Benchmark configuration for FREYA SLAM pipeline
+# Sequences: dataset-room1_512_16, dataset-room2_512_16, dataset-corridor1_512_16
+
+[freya]
+device = "cuda:0"
+batch_size = 1
+log_level = "INFO"
+
+[freya.trajectory]
+bspline_knot_spacing = 0.1
+bspline_order = 4
+bundle_adj_max_iter = 30  # Paper: N_t = 30
+
+[freya.gaussian]
+max_gaussians = 23000000  # 23M gaussians → ~18.8GB params+Adam on L4 (23GB)
+target_vram_gb = 20.0     # auto-size if set; overrides max_gaussians
+prune_opacity_threshold = 0.005
+sh_degree = 3
+
+[freya.depth]
+model = "depth_anything_v2_small"
+model_path = "/mnt/forge-data/models/depth-anything--Depth-Anything-V2-Small"
+
+[freya.rendering]
+width = 512   # TUM VI resolution
+height = 512
+tile_size = 16
+
+[freya.sensors]
+sync_tolerance_ms = 50.0
+lidar_rate_hz = 0.0      # No LiDAR in TUM VI — pseudo-LiDAR from stereo
+camera_rate_hz = 20.0     # TUM VI cam rate
+imu_rate_hz = 200.0       # TUM VI IMU rate
+
+[freya.paths]
+dataset_root = "/mnt/forge-data/datasets/tum"
+calibration_dir = "configs/calibration"
+results_dir = "/mnt/artifacts-datai/reports/project_freya"
+checkpoint_dir = "/mnt/artifacts-datai/checkpoints/project_freya"
+log_dir = "/mnt/artifacts-datai/logs/project_freya"
+tensorboard_dir = "/mnt/artifacts-datai/tensorboard/project_freya"
+export_dir = "/mnt/artifacts-datai/exports/project_freya"
+model_dir = "/mnt/forge-data/models"
+
+[freya.paper]
+# Paper hyperparameters (Section X-A)
+N_t = 30          # Camera pose optimization steps
+loss_lambda = 0.2 # Loss weighting
+xi = 0.005        # Pruning threshold
+tau = 0.99        # Map expansion threshold
+K = 100           # Selected keyframes
+epsilon_1 = 0.1   # Depth regularization min (meters)
+epsilon_2 = 50.0  # Depth regularization max (meters)

onnx/freya_depth_v1.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ff5e95e22c28d8cbffe3d223c5d099a0c4acd662d4a270d6837c95483a1f20c3
+size 1114

pytorch/freya_v1.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:48e91f151633cd2f67c373d1b35e229221ca64561d5876202b35cde0d93bdc1e
+size 135769

pytorch/freya_v1.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cbc519d61e4175ce7635aa805d0e9e2d2017f4338a105cf3cdbc78a6843e6687
+size 58392