Fix GPU model + add training/eval repo links

README.md

@@ -21,6 +21,16 @@ From-scratch [Diffusion Policy](https://arxiv.org/abs/2303.04137) fine-tune on
 for the LeIsaac `LeIsaac-SO101-PickOrange-v0` task in Isaac Sim (SO-101 leader arm,
 front + wrist cameras, three oranges → plate).
 
+## Project links
+
+- **Training repo**: [`vitorcen/LeIsaac`](https://github.com/vitorcen/LeIsaac) —
+  fork of LightwheelAI/leisaac with the reusable fine-tune scaffold
+  (`scripts/finetune/lerobot_finetune.sh`), datasets workflow, and design docs
+  used to produce this checkpoint.
+- **Evaluation harness**: [`vitorcen/isaaclab-experience`](https://github.com/vitorcen/isaaclab-experience) —
+  the Isaac Sim host project (LeIsaac live preview + LeRobot async-inference
+  policy_server launcher) used to run the eval rounds reported below.
+
 ## Headline result
 
 | Eval budget | Episodes | Success rate |
@@ -50,7 +60,7 @@ Compared to other policies trained on the same 60 ep × 36 k frame dataset:
 - **Scheduler**: cosine with 500 warmup steps
 - **Batch**: 32 per device
 - **Steps**: 100,000 (resume @ 60k saved ckpt → continued to 100k under same config)
-- **Time on 1× RTX 3090 24 GB**: ~8 h total, ~2.5 step/s steady-state
+- **Time on 1× RTX 4090 24 GB**: ~8 h total, ~2.5 step/s steady-state
 - **Video backend**: pyav (torchcodec + 4 workers segfaults on long runs)
 - **Wandb**: disabled