LIBERO-Spatial checkpoints: world model + diffusion policy

README.md

@@ -0,0 +1,24 @@
+---
+license: other
+tags:
+- robotics
+- world-model
+- diffusion-policy
+- libero
+---
+
+# 4D World Models for Robotic Manipulation -- checkpoints
+
+Checkpoints for the LIBERO-Spatial results of the accompanying code repository.
+
+| File | Model | Parameters |
+|---|---|---|
+| `world_model/best.pt` | 4D particle-flow world model (EMA weights) | 35.1 M |
+| `policy_libero_spatial/best_action_policy_ema.pt` | flow- and physics-conditioned diffusion policy (EMA weights) | 156.2 M |
+| `policy_libero_spatial/interface.json` | policy conditioning interface (read by the evaluator) | -- |
+
+LIBERO-Spatial closed loop: 47 / 50 successes (94 %), 10 tasks x the first 5 official initial states.
+
+```bash
+hf download LIN-linlin/4D-World-Models-Robotic-Manipulation --local-dir checkpoints
+```

policy_libero_spatial/best_action_policy_ema.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f7ab3f69316037983f91db6da6f77a50b6eb7a3ddff6c9b0a1e230e8b907554b
+size 625291324

policy_libero_spatial/interface.json

@@ -0,0 +1,13 @@
+{
+ "phys_in_dim": 384,
+ "phys_mode": "film",
+ "plan_in_dim": 13,
+ "d_plan": 128,
+ "global_cond_dim": 1408,
+ "phys_source": "phys_latent",
+ "phys_layout": "latent",
+ "plan_layout": "v1",
+ "n_points": 64,
+ "wrist_drop": 0.2,
+ "image_drop": 0.1
+}

world_model/best.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:476d6d6555612cdd83fbfa44f5977698eb2e9dfa09d61f9a1a27bc429a348fd6
+size 140341813