Update model card - remove author/date info

README.md

@@ -11,135 +11,63 @@ tags:
 - tensorrt
 datasets:
 - libero
-metrics:
-- accuracy
 pipeline_tag: robotics
 library_name: pytorch
 ---
 
-# Turbo-Pi0.5: High-Performance VLA Model on NVIDIA Jetson Thor
+# Turbo-Pi0.5
 
-**14.7x Speedup** | **20.6 Hz Inference** | **98% LIBERO Spatial Accuracy**
+**14.7x faster Pi0.5 VLA model optimized for NVIDIA Jetson Thor**
 
-Turbo-Pi0.5 is an optimized implementation of the Pi0.5 Vision-Language-Action (VLA) model for NVIDIA Jetson Thor platform, achieving real-time robot control at 20.6 Hz.
-
-## Model Description
-
-This model is based on the [Physical Intelligence Pi0](https://www.physicalintelligence.company/blog/pi0) architecture, optimized for edge deployment on NVIDIA Jetson Thor with the following techniques:
-
-- **KV Cache**: Reuse prefix K,V across denoising steps
-- **TensorRT Export**: ONNX to TensorRT FP16 engines
-- **Reduced Denoising Steps**: 10 to 3 steps with minimal accuracy loss
-- **Dual Stream Pipeline**: Vision-Action parallel execution
-
-## Performance
-
-| Metric | Baseline | Turbo-Pi0.5 | Improvement |
-|--------|----------|-------------|-------------|
-| **Throughput** | 1.4 Hz | **20.6 Hz** | **14.7x** |
-| **Latency** | 714 ms | **48.5 ms** | **14.7x** |
-| **LIBERO Spatial** | - | **98%** | - |
-
-### TensorRT Performance (NVIDIA Jetson Thor)
-
-| Denoising Steps | Latency | Throughput |
-|-----------------|---------|------------|
-| 10 steps | 131.7 ms | 7.6 Hz |
-| 5 steps | 72.3 ms | 13.8 Hz |
-| **3 steps** | **48.5 ms** | **20.6 Hz** |
-| 2 steps | 36.7 ms | 27.3 Hz |
-
-## LIBERO Benchmark Results
-
-### libero_spatial (10 tasks, 10 trials each)
-
-| Task | Success Rate |
-|------|--------------|
-| pick_up_black_bowl_between_plate_and_ramekin | 10/10 (100%) |
-| pick_up_black_bowl_next_to_ramekin | 10/10 (100%) |
-| pick_up_black_bowl_from_table_center | 10/10 (100%) |
-| pick_up_black_bowl_next_to_cookie_box | 10/10 (100%) |
-| pick_up_black_bowl_in_top_drawer | 10/10 (100%) |
-| pick_up_black_bowl_on_ramekin | 8/10 (80%) |
-| pick_up_black_bowl_on_cookie_box | 10/10 (100%) |
-| pick_up_black_bowl_on_stove | 10/10 (100%) |
-| pick_up_black_bowl_next_to_plate | 10/10 (100%) |
-| pick_up_black_bowl_on_wooden_cabinet | 10/10 (100%) |
-| **Total** | **98/100 (98%)** |
+| Metric | Before | After |
+|--------|--------|-------|
+| Inference Speed | 1.4 Hz | **20.6 Hz** |
+| Latency | 714 ms | **48.5 ms** |
+| LIBERO Spatial | - | **98%** |
 
 ## Usage
 
-### Installation
-
 ```bash
-# Clone repository
-git clone https://github.com/LiangSu8899/TurboPi.git
-cd TurboPi/openpi
-
-# Install dependencies
-pip install -e .
-
 # Download model
-huggingface-cli download liangsu9988/Turbo-Pi0.5 --local-dir ~/.cache/openpi/checkpoints/pi05_libero
-```
+huggingface-cli download liangsu9988/Turbo-Pi0.5 \
+    --local-dir ~/.cache/openpi/checkpoints/pi05_libero
 
-### Inference
-
-```bash
-# Start policy server
-python scripts/serve_policy.py \
-    --env=LIBERO \
-    --port=8000 \
-    policy:checkpoint \
-    --policy.config=pi05_libero \
-    --policy.dir=~/.cache/openpi/checkpoints/pi05_libero
+# Run inference server
+python scripts/serve_policy.py --env=LIBERO --port=8000 \
+    policy:checkpoint --policy.config=pi05_libero
 ```
 
-### Python API
+## Architecture
 
-```python
-from openpi_client import WebsocketClientPolicy
+- **Vision**: SigLIP-SO400M
+- **Language**: Gemma 2B
+- **Action**: Gemma 300M Expert
 
-# Connect to policy server
-policy = WebsocketClientPolicy(host="localhost", port=8000)
+## Optimizations
 
-# Get action from observation
-action = policy.get_action({
-    "images": {"cam_high": image, "cam_left_wrist": left_img, "cam_right_wrist": right_img},
-    "state": robot_state,
-    "prompt": "pick up the black bowl"
-})
-```
+- KV Cache for efficient denoising
+- TensorRT FP16 acceleration
+- Reduced denoising steps (10 → 3)
 
-## Model Architecture
+## LIBERO Spatial Results (98%)
 
-- **Vision Encoder**: SigLIP-SO400M (400M parameters)
-- **Language Model**: Gemma 2B
-- **Action Expert**: Gemma 300M
-- **Total Parameters**: ~2.7B
-
-## Hardware Requirements
-
-- NVIDIA Jetson Thor (recommended) or GPU with 8GB+ VRAM
-- JetPack 7.1+ for Jetson
-- CUDA 12.0+
-
-## Citation
-
-```bibtex
-@misc{turbopi05,
-  title={Turbo-Pi0.5: High-Performance VLA Model on NVIDIA Jetson Thor},
-  author={Liang Su},
-  year={2026},
-  url={https://github.com/LiangSu8899/TurboPi}
-}
-```
+| Task | Success |
+|------|---------|
+| pick_up_black_bowl_between_plate_and_ramekin | 100% |
+| pick_up_black_bowl_next_to_ramekin | 100% |
+| pick_up_black_bowl_from_table_center | 100% |
+| pick_up_black_bowl_next_to_cookie_box | 100% |
+| pick_up_black_bowl_in_top_drawer | 100% |
+| pick_up_black_bowl_on_ramekin | 80% |
+| pick_up_black_bowl_on_cookie_box | 100% |
+| pick_up_black_bowl_on_stove | 100% |
+| pick_up_black_bowl_next_to_plate | 100% |
+| pick_up_black_bowl_on_wooden_cabinet | 100% |
 
-## Acknowledgments
+## Links
 
-- [Physical Intelligence](https://www.physicalintelligence.company/) for the original Pi0 model
-- [OpenPi](https://github.com/Physical-Intelligence/openpi) for the open-source implementation
-- NVIDIA for Jetson Thor platform and TensorRT tools
+- **Code**: [GitHub](https://github.com/LiangSu8899/TurboPi)
+- **Base**: [OpenPi](https://github.com/Physical-Intelligence/openpi)
 
 ## License