Update README.md
Browse files
README.md
CHANGED
|
@@ -1,3 +1,61 @@
|
|
| 1 |
---
|
| 2 |
license: cc-by-nc-sa-4.0
|
| 3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
license: cc-by-nc-sa-4.0
|
| 3 |
---
|
| 4 |
+
|
| 5 |
+
# AlpaSim E2E Challenge 2026 — NuPlan / MTGS Track Data
|
| 6 |
+
|
| 7 |
+
This dataset contains the official evaluation assets for the **NuPlan / MTGS Track** of the [AlpaSim E2E Closed-Loop Challenge 2026](https://huggingface.co/spaces/nvidia/AlpasimE2EClosedLoopChallenge2026).
|
| 8 |
+
|
| 9 |
+
These resources are managed exclusively by the **trusted evaluator**. Contestants do **not** need to package or depend on them in their submitted driver images.
|
| 10 |
+
|
| 11 |
+
---
|
| 12 |
+
|
| 13 |
+
## Dataset Structure
|
| 14 |
+
|
| 15 |
+
```
|
| 16 |
+
MTGS_asset/
|
| 17 |
+
└── navtest/
|
| 18 |
+
├── assets/
|
| 19 |
+
│ ├── part001.tar.gz
|
| 20 |
+
│ ├── ...
|
| 21 |
+
│ └── part015.tar.gz # MTGS neural rendering assets (~456 GB total)
|
| 22 |
+
└── configs.tar.gz # navtest scene configurations
|
| 23 |
+
|
| 24 |
+
trajdata_cache/
|
| 25 |
+
└── nuplan_test.tar.gz # trajdata cache for nuPlan test scenes
|
| 26 |
+
```
|
| 27 |
+
|
| 28 |
+
### `MTGS_asset/navtest/`
|
| 29 |
+
|
| 30 |
+
Neural rendering assets used by the [MTGS](https://github.com/OpenDriveLab/MTGS) renderer during closed-loop evaluation. The evaluator mounts these assets into the simulator stack; contestant containers do not have direct access to them.
|
| 31 |
+
|
| 32 |
+
### `trajdata_cache/nuplan_test.tar.gz`
|
| 33 |
+
|
| 34 |
+
Pre-built [trajdata](https://github.com/NVlabs/trajdata) cache for the nuPlan test split. The evaluator uses this cache to initialize and drive nuPlan scenes during evaluation rollouts.
|
| 35 |
+
|
| 36 |
+
---
|
| 37 |
+
|
| 38 |
+
## How Evaluation Works
|
| 39 |
+
|
| 40 |
+
1. The evaluator starts the full AlpaSim simulator stack (physics, traffic, MTGS renderer) using the assets in this dataset.
|
| 41 |
+
2. The contestant's driver container receives rendered camera images and scene state over gRPC and must respond with control commands within **0.1 s** per `Drive` call.
|
| 42 |
+
3. Results are aggregated and posted to the leaderboard.
|
| 43 |
+
|
| 44 |
+
The official evaluation preset is `+e2e_challenge=ec2` (`topology=8gpu_36rollouts`): 12 contestant replicas across GPUs 4–7, each handling 3 concurrent rollouts.
|
| 45 |
+
|
| 46 |
+
---
|
| 47 |
+
|
| 48 |
+
## Contestant Submission Requirements
|
| 49 |
+
|
| 50 |
+
- Implement `egodriver.EgodriverService` (see `src/grpc/alpasim_grpc/v0/egodriver.proto`).
|
| 51 |
+
- Image size ≤ **40 GiB**; do not bundle nuPlan data, trajdata caches, navtest configs, or MTGS assets.
|
| 52 |
+
- Outbound network access is blocked during evaluation; the root filesystem is read-only.
|
| 53 |
+
- Writable scratch space: `/tmp` (2 GiB), `/run` (64 MiB).
|
| 54 |
+
|
| 55 |
+
For full instructions see the [AlpaSim E2E Challenge README](https://huggingface.co/spaces/nvidia/AlpasimE2EClosedLoopChallenge2026).
|
| 56 |
+
|
| 57 |
+
---
|
| 58 |
+
|
| 59 |
+
## License
|
| 60 |
+
|
| 61 |
+
This dataset is released under [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).
|