Downloads · 30 days
0
hwanghwijun/libero10-hybrid-v3-half-mix-best-checkpoints
libero10-hybrid-v3-half-mix-best-checkpoints is a robotics model from hwanghwijun. Use it for the robotics task on the model card, and read the license before you ship it in a product. The card lists the license as mit.
Control-condition checkpoints for a trajectory-retrieval ablation on LIBERO-10.
Downloads · 30 days
0
Access
Public
Updated Aug 26, 2026
Repo size
4.2 GB
Likes
0
Public
Click a slice to open those files.
.pth4.2 GB · 100%
From the Hugging Face model README
Control-condition checkpoints for a trajectory-retrieval ablation on LIBERO-10.
A naive mixing baseline, not a proposed method. For each task, given the
S-DTW retrieval result R_S (N matches) and the standalone VP-Bresenham result
R_B (N matches), with H = N // 2:
N - H lowest-cost matches of R_SH lowest-cost matches of R_BThe two retrieval results are never compared against each other; each is ranked by its own accumulated S-DTW cost. There is no per-segment matching and no cosine similarity computation.
30 checkpoints = 10 tasks x 3 seeds (1234 / 42 / 4325), one per run, at that run's
best-rollout epoch. Layout: <task>/seed<seed>/model_epoch_<E>.pth.
| Task | Hybrid v1 best-K | Hybrid v2 best-K | v3 |
|---|---|---|---|
| stove_moka | 0.810 | 0.853 | 0.773 |
| soup_cheese | 0.397 | 0.347 | 0.323 |
| mug_mug | 0.457 | 0.537 | 0.430 |
| bowl_cabinet | 0.980 | 0.983 | 0.970 |
| book_caddy | 0.703 | 0.700 | 0.710 |
| mug_microwave | 0.320 | 0.233 | 0.213 |
| moka_moka | 0.000 | 0.000 | 0.000 |
| soup_sauce | 0.523 | 0.537 | 0.270 |
| mug_pudding | 0.217 | 0.237 | 0.193 |
| cream_butter | 0.603 | 0.527 | 0.333 |
| Mean | 0.501 | 0.495 | 0.422 |
v3 is lower than the v1 best-K setting on 8 of 10 tasks (mean gap -0.079) and lower
than the v2 best-K setting on 8 of 10 (mean gap -0.074). book_caddy is the single
counterexample (+0.007); moka_moka is 0.000 for every method.
Note that "best-K" is a per-task oracle selection over K in {5,10,15,20,25,30}, whereas v3 is a single fixed configuration.
BC-Transformer-GMM (5 modes, frame_stack 5, seq_length 5, batch 32), 300 epochs x 200
gradient steps, rollout every 50 epochs with 100 episodes, horizon 300,
terminate-on-success. Retrieval queries were fixed to demo_45, demo_15, demo_1, demo_7, demo_24 for every task and method.
Environment: Python 3.10.20, torch 2.5.1+cu121, mujoco 3.6.0, robosuite 1.4.0, robomimic 0.3.0, RTX 4090.