environment string | basis list | basis_weights list | temperature float64 | level_scales list | perturb_scales list | std_normalize bool | student_policy string | nsample int64 | hsample int64 | hnode int64 | ndiffuse int64 | seconds float64 | queries int64 | env_steps int64 | mean_reward float64 | falls int64 | num_envs int64 | repeats int64 | levels int64 | queries_per_level int64 | steps_per_call int64 | shards list |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
unitree_go2_gait | [
"e0",
"e1",
"e2",
"e3"
] | [
[
1,
0,
0,
0
],
[
0,
1,
0,
0
],
[
0,
0,
1,
0
],
[
0,
0,
0,
1
]
] | 0.15 | [
1,
1
] | [
0.25,
0.5,
1,
1.5
] | false | csm_runs/gait-fit-v2/clouds-fit-20260915-023426/policy.pkl | 2,049 | 32 | 8 | 2 | 28,834.511801 | 123,200 | 12,000 | -1.545349 | 31 | 8 | 8 | 2 | 5 | 25 | [
"clouds_0000.npz",
"clouds_0001.npz",
"clouds_0002.npz",
"clouds_0003.npz",
"clouds_0004.npz",
"clouds_0005.npz"
] |
unitree_go2_gait | [
"e0",
"e1",
"e2",
"e3"
] | [
[
1,
0,
0,
0
],
[
0,
1,
0,
0
],
[
0,
0,
1,
0
],
[
0,
0,
0,
1
]
] | 0.15 | [
1,
1
] | [
0.25,
0.5,
1,
1.5
] | false | null | 2,049 | 32 | 8 | 2 | 34,604.947658 | 164,480 | 16,000 | -0.9394 | 1 | 8 | 8 | 2 | 5 | 25 | [
"clouds_0000.npz",
"clouds_0001.npz",
"clouds_0002.npz",
"clouds_0003.npz",
"clouds_0004.npz",
"clouds_0005.npz",
"clouds_0006.npz",
"clouds_0007.npz"
] |
unitree_go2_push_recover | [
"boost0",
"boost1",
"boost2",
"boost3"
] | [
[
0.8660253882408142,
0.28867512941360474,
0.28867512941360474,
0.28867512941360474
],
[
0.28867512941360474,
0.8660253882408142,
0.28867512941360474,
0.28867512941360474
],
[
0.28867512941360474,
0.28867512941360474,
0.8660253882408142,
0.28867512941360474
... | 0.25 | [
3.175,
1
] | [
0.25,
0.5,
1,
1.5
] | false | null | 2,049 | 24 | 6 | 2 | 26,577.586181 | 422,080 | 41,088 | -0.467755 | 0 | 8 | 2 | 2 | 5 | 25 | [
"clouds_0000.npz",
"clouds_0001.npz",
"clouds_0002.npz",
"clouds_0003.npz",
"clouds_0004.npz",
"clouds_0005.npz",
"clouds_0006.npz",
"clouds_0007.npz",
"clouds_0008.npz",
"clouds_0009.npz",
"clouds_0010.npz",
"clouds_0011.npz",
"clouds_0012.npz",
"clouds_0013.npz",
"clouds_0014.npz",
"... |
unitree_go2_walk | [
"boost0",
"boost1",
"boost2"
] | [
[
0.9045340418815613,
0.30151134729385376,
0.30151134729385376
],
[
0.30151134729385376,
0.9045340418815613,
0.30151134729385376
],
[
0.30151134729385376,
0.30151134729385376,
0.9045340418815613
]
] | 0.02 | [
2.625,
1
] | [
0.25,
0.5,
1,
1.5
] | false | null | 2,049 | 16 | 4 | 2 | 19,605.97344 | 821,440 | 80,000 | -0.054135 | 0 | 8 | 2 | 2 | 5 | 25 | [
"clouds_0000.npz",
"clouds_0001.npz",
"clouds_0002.npz",
"clouds_0003.npz",
"clouds_0004.npz",
"clouds_0005.npz",
"clouds_0006.npz",
"clouds_0007.npz",
"clouds_0008.npz",
"clouds_0009.npz",
"clouds_0010.npz",
"clouds_0011.npz",
"clouds_0012.npz",
"clouds_0013.npz",
"clouds_0014.npz",
"... |
DIAL-MPC CSM sample clouds — Unitree Go2
The raw training data behind the three released Go2 controllers of
joonseoJ/dial-mpc (release models-v1,
branch csm/compositional-score-matching). Each file is a cloud collection:
for every query plan, the per-row horizon cost of every one of DIAL-MPC's 2049
sampled action sequences, plus the RNG keys that regenerate those sequences.
Any label -- weight, temperature, per-level temperature profile, number of
repeats -- is re-derived from these by a softmax with no physics
(csm/cloud_data.py).
| Folder | Task | Size | Queries | Model it produced |
|---|---|---|---|---|
pr_collect_v2/ |
push recovery (4 rows) | 27 GB | 422,080 | pr-v2-fit/clouds-fit-20260908-202201 |
walk_collect_v5/ |
walking (tracking / stability / gait) | 39 GB | 821,440 | clouds-fit-20260905-083613 |
gait_v2/ + gait_d1/ |
gait styles (walk / trot / pace / bound), base + one DAgger round | 72 GB | 287,030 | gait-fit-d1/clouds-fit-20260915-115906 |
Each folder holds clouds_*.npz shards and a collection.json with the
settings the collection ran under. Shard fields (DialCloudData):
u (Q, H, 12)query plans, normalized node coordinates in [-1, 1]factor (Q,),level (Q,)annealing factor and level of each queryobs (Q, obs_dim),qpos,qvel,stepthe state the query was posed atsample_rng (Q, R, 2)one key per repeat; regenerates the 2048 samplesterms (Q, R, 2049, k)horizon-mean reward of every sample per row (last sample = the query itself)noise (Q, H)per-node sampling sigma
Use
git clone -b csm/compositional-score-matching https://github.com/joonseoJ/dial-mpc
cd dial-mpc && uv venv -p 3.10 && source .venv/bin/activate && uv pip install -e .
hf download joonseo-jang/dial-mpc-csm-clouds --repo-type dataset --local-dir csm_runs
# refit the walking fields exactly as released
python -m csm.fit_from_clouds --clouds csm_runs/walk_collect_v5 \
--example unitree_go2_trot_csm --basis boost0 boost1 boost2 \
--temperature 0.020 --level-scales 2.625 1.0 \
--hidden 768,768,768 --train-iters 300000 --batch-size 512
The shards are only valid with the observation convention of the code they
were collected under (walking: 54-dim, translation-invariant). Changing
_get_obs invalidates them.
- Downloads last month
- 120