fs int64 5 5 | action_raw_dim int64 2 22 | img_size int64 224 224 | encoder_size stringclasses 1
value | encoder_backbone stringclasses 1
value | encoder_ckpt null | z_dim int64 192 192 | proj_hidden int64 384 384 | d_model int64 192 192 | n_heads int64 4 4 | depth int64 4 4 | dropout float64 0.1 0.1 | n_flow_steps int64 8 8 | num_actions_pred int64 10 10 | num_states_pred int64 1 1 | action_head stringclasses 1
value | state_head stringclasses 1
value | num_views int64 1 1 | views listlengths 1 1 | goal_views listlengths 1 1 | sigreg_proj_dim int64 0 0 | action_mean listlengths 2 22 | action_std listlengths 2 22 | input_norm stringclasses 1
value | input_norm_mean listlengths 3 3 | input_norm_std listlengths 3 3 | H_max int64 10 10 | history_len int64 2 2 | history_stride int64 5 5 | goal_type stringclasses 1
value | goal_view_indices listlengths 1 1 | sigreg_grouped bool 1
class | flow_alpha_act float64 1 1 | flow_alpha_state float64 1 1 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
5 | 5 | 224 | tiny | resnet18dp | null | 192 | 384 | 192 | 4 | 4 | 0.1 | 8 | 10 | 1 | flow | mse | 1 | [
"pixels"
] | [
"pixels"
] | 0 | [
0.010884694755077362,
-0.0031414327677339315,
0.002646582666784525,
0.00042392851901240647,
0.15925255417823792
] | [
0.28941991925239563,
0.39371708035469055,
0.6431366801261902,
0.3928017318248749,
0.25030744075775146
] | imagenet | [
0.48500001430511475,
0.4560000002384186,
0.4059999883174896
] | [
0.2290000021457672,
0.2240000069141388,
0.22499999403953552
] | 10 | 2 | 5 | sampled | [
0
] | true | 1 | 1 |
5 | 22 | 224 | tiny | resnet18dp | null | 192 | 384 | 192 | 4 | 4 | 0.1 | 8 | 10 | 1 | flow | mse | 1 | [
"pixels"
] | [
"pixels"
] | 0 | [
0.14831849932670593,
-0.11819789558649063,
-0.06229739636182785,
-0.019640153273940086,
0.06845644861459732,
0.01653614267706871,
0.3483588397502899,
0.4423060417175293,
0.46674907207489014,
0.39561882615089417,
0.0481521375477314,
-0.03765391930937767,
-0.005276348441839218,
-0.1017925441... | [
0.5038843750953674,
0.5541285276412964,
0.5030153393745422,
0.1473264843225479,
0.28025931119918823,
0.11758679151535034,
0.5413413047790527,
0.5647051930427551,
0.572503924369812,
0.5353013873100281,
0.12254844605922699,
0.4814732074737549,
0.5450199842453003,
0.49891766905784607,
0.104... | imagenet | [
0.48500001430511475,
0.4560000002384186,
0.4059999883174896
] | [
0.2290000021457672,
0.2240000069141388,
0.22499999403953552
] | 10 | 2 | 5 | sampled | [
0
] | true | 1 | 1 |
5 | 2 | 224 | tiny | resnet18dp | null | 192 | 384 | 192 | 4 | 4 | 0.1 | 8 | 10 | 1 | flow | mse | 1 | [
"pixels"
] | [
"pixels"
] | 0 | [
-0.007812565192580223,
0.006860686466097832
] | [
0.2084674835205078,
0.20674866437911987
] | imagenet | [
0.48500001430511475,
0.4560000002384186,
0.4059999883174896
] | [
0.2290000021457672,
0.2240000069141388,
0.22499999403953552
] | 10 | 2 | 5 | sampled | [
0
] | true | 1 | 1 |
5 | 5 | 224 | tiny | resnet18dp | null | 192 | 384 | 192 | 4 | 4 | 0.1 | 8 | 10 | 1 | flow | mse | 1 | [
"pixels"
] | [
"pixels"
] | 0 | [
0.020879777148365974,
-0.0052217054180800915,
0.045017410069704056,
0.02976837009191513,
0.054552823305130005
] | [
0.4004173278808594,
0.42906442284584045,
0.37829679250717163,
0.6607723236083984,
0.12185002863407135
] | imagenet | [
0.48500001430511475,
0.4560000002384186,
0.4059999883174896
] | [
0.2290000021457672,
0.2240000069141388,
0.22499999403953552
] | 10 | 2 | 5 | sampled | [
0
] | true | 1 | 1 |
5 | 5 | 224 | tiny | resnet18dp | null | 192 | 384 | 192 | 4 | 4 | 0.1 | 8 | 10 | 1 | flow | mse | 1 | [
"pixels"
] | [
"pixels"
] | 0 | [
0.009110543876886368,
0.005368020851165056,
0.02910570614039898,
0.08754989504814148,
0.09633138030767441
] | [
0.4143800139427185,
0.43083059787750244,
0.5773980021476746,
0.5810631513595581,
0.25676199793815613
] | imagenet | [
0.48500001430511475,
0.4560000002384186,
0.4059999883174896
] | [
0.2290000021457672,
0.2240000069141388,
0.22499999403953552
] | 10 | 2 | 5 | sampled | [
0
] | true | 1 | 1 |
YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
LeWAM standard (sig) checkpoints — 2026-10-05
Best-validation checkpoints for the standard LeWAM recipe on five cells, plus the full-dataset PushT checkpoint the main PushT table quotes.
Recipe (identical for every *_sig run)
stride-1 data path, SIGReg w_reg=0.04 (proj_dim=0, per-time), H_max=10, 50 epochs, 5 warmup epochs,
batch 128, lr 1.5e-4 cosine, seed 42, frameskip 5, 224 px.
Architecture (all cells): z_dim=192, d_model=192, depth 4, 4 heads, n_flow_steps=8,
num_actions_pred=10, num_states_pred=1, policy_history_len=2, encoder resnet18dp (tiny).
Code: lewam-jointflow@2d52888.
Contents
Each directory holds jointflow_best.pt (best validation) and jointflow_config.json (architecture plus the
action z-score stats and frameskip needed to un-normalise predictions).
| directory | cell | source run | action dim | training data |
|---|---|---|---|---|
scene/ |
OGBench scene | gr_scene_sig |
5 | scene_expert.h5, full |
puzzle/ |
OGBench puzzle 3x3 | gr_puzzle_sig |
5 | puzzle_3x3_expert.h5, full |
cube/ |
OGBench cube | gr_cube_sig |
5 | cube.h5, full |
drawer/ |
DexMimicGen TwoArmDrawerCleanup | gr_drawer_sig |
22 | drawer_cleanup_fixed.h5, full |
pusht/ |
Push-T | gr_pusht_ep16 |
2 | pusht.h5, full, epoch 16 |
Note on Push-T. This is the full-dataset checkpoint the main Push-T results table quotes. The Push-T
ablation family (sig_sub, nosig_sub, mse_sub, msenosig_sub) was trained on a 77% episode subset and is
deliberately not in this repo: those arms are only comparable to each other, not to the full-dataset runs.
Goal-reaching success (50 (episode, start) pairs, seeds 42/0/1, mean ± sd)
Protocol differs by cell and always has: scene / puzzle / cube / Push-T use goal offset 25, budget 50, plan horizon 5; drawer uses goal offset 50, budget 100, plan horizon 10. Numbers from different protocols are not comparable.
| cell | gc (reactive) | plan (best-of-32) | grad_allk + TR |
|---|---|---|---|
| scene | 90.7 ± 5.8 | 96.0 ± 2.0 | 97.3 ± 1.2 |
| puzzle | 76.7 ± 8.1 | 76.7 ± 10.1 | 78.7 ± 9.5 |
| cube | 99.3 ± 1.2 | 100.0 ± 0.0 | 100.0 ± 0.0 |
| drawer | 64.0 ± 3.5 | 64.0 ± 8.7 | 60.0 ± 7.2 |
| pusht | 85.3 ± 5.0 | 95.3 ± 2.3 | 98.7 ± 1.2 |
Success rule: object-position distance below the cell's object threshold and end-effector distance below its eef threshold, "ever reached" within the budget. drawer additionally latches the environment's own task predicate.
Push-T latent probing (ridge R², effective rank = spectral entropy)
Action (IDM) is reported as z_t-only (z_t and z_t+K).
| model | SIGReg | eff. rank | Object | Proprio | Action (IDM) |
|---|---|---|---|---|---|
LeWAM (flow), sig_sub |
Yes | 43.92 | 0.88 | 0.86 | 0.38 (0.77) |
LeWAM (flow), nosig_sub |
No | 10.94 | 0.88 | 0.85 | 0.34 (0.72) |
LeWAM (MSE), mse_sub |
Yes | 45.84 | 0.91 | 0.84 | 0.34 (0.76) |
LeWAM (MSE), msenosig_sub |
No | 6.96 | 0.66 | 0.86 | 0.36 (0.74) |
| LeWM, official weights | Yes | 94.21 | 0.95 | 0.80 | 0.29 (0.71) |
| LeWM, official code, no SIGReg | No | 16.23 | 0.00 | 0.00 | 0.00 (0.00) |
None of these probing rows is the pusht/ checkpoint in this repo — they are the 77%-subset ablation family,
listed only so the SIGReg and action-head effects are on record. Do not mix them with pusht/ in one table.
Checksums (jointflow_best.pt)
| directory | md5 (first 12) | size |
|---|---|---|
| scene | e689d11b8ad0 | 61 MB |
| puzzle | 7282f3b01b03 | 61 MB |
| cube | b798cfaec476 | 61 MB |
| drawer | 081c8dadae89 | 61 MB |
| pusht | a9c31ad24f50 | 61 MB |
Machine-readable performance
checkpoint_performance.csv — one row per (checkpoint, planner), with the protocol that produced it:
checkpoint_dir, cell, action_dim, training_data, md5_best_pt, goal_offset_steps, eval_budget, plan_horizon_blocks, planner, success_mean, success_sd, n_seeds, per_seed_42_0_1
Always read goal_offset_steps / eval_budget / plan_horizon_blocks before comparing two rows: drawer runs
50 / 100 / 10 while every other cell runs 25 / 50 / 5, and those numbers are not comparable.
v1 checkpoint format
The five checkpoints and JSON configs have been migrated to the v1 model schema.
Directory and file names remain unchanged: jointflow_best.pt now contains a v1
state dict, and jointflow_config.json contains v1 model keys. The original
training provenance and success-rate tables above are historical results.
Every retained tensor is exactly equal to its source tensor after a one-to-one
key rename. Strict loading, attention masks, encoder output, supplied-goal policy
ODE output, and dynamics output were checked against lewam-jointflow@2d52888.
The removed null_goal tensor was verified to be all zero in every checkpoint;
these runs used zero goal dropout. Goal-absent/dropout behavior follows v1.
To use the normal v1 evaluation loader, download the unchanged remote filenames
as lewam_best.pt and lewam_config.json in a local run directory under
$STABLEWM_HOME/checkpoints/. No optimizer state is included.
- Downloads last month
- 95