MolmoAct2-SO101-VTP_toponly_p20_Init
Step-0 checkpoint: compression applied, training not started.
Grid Sampler vision-token pruning on the overhead (top) camera only, at 20.4% — 156 of 196 tokens kept. The wrist camera is left untouched: it keeps its full 196-token pooled grid and does not pass through a sampler at all.
The Grid Sampler weights are randomly initialized. This is a training seed, not something to evaluate.
Why top-only, and why "untouched" is not the same as K=196
The two views carry different kinds of information — the overhead view gives global layout (where the cube and box are), the wrist view gives fine detail near the gripper. Pruning one and not the other isolates which view tolerates compression, which is directly actionable for on-device deployment.
A camera marked "not pruned" here bypasses the sampler entirely. Asking for
K=196 instead would still run the ActiveTokenSampler: it predicts 196
coordinates, bilinearly resamples the feature map at them, and adds a coordinate
embedding. Same token count, but not an identity — and it adds randomly initialized
parameters to a path that needed none. This checkpoint takes the bypass, verified by
asserting the wrist tokens come out torch.equal to the pooled grid that went in.
| Baseline | This variant | |
|---|---|---|
| top camera tokens | 196 | 156 (learned sample points) |
| wrist camera tokens | 196 | 196 (unchanged, sampler bypassed) |
| image tokens total | 392 | 352 (10.2% fewer) |
| Grid Sampler modules | — | 1 (top only) |
| New parameters | — | 2,129,976 (randomly initialized) |
| Action expert / flow matching | unchanged | unchanged |
Where this sits
| Variant | top | wrist | image tokens | Grid Sampler modules |
|---|---|---|---|---|
| no pruning | 196 | 196 | 392 | 0 |
| toponly p20 | 156 | 196 | 352 | 1 |
| toponly p40 | 117 | 196 | 313 | 1 |
| toponly p60 | 78 | 196 | 274 | 1 |
| VTP (both cams, K=16) | 16 | 16 | 32 | 1 (shared) |
This variant: top 156 tokens (20.4% pruned), wrist untouched.
Honest expectation
The premise "the wrist view must stay intact" is not supported by the evidence that exists. On LIBERO the Grid Sampler was applied uniformly to both cameras (agentview and wrist) and K=16 — 91.8% pruning on both — held parity with the unpruned control (97.1 vs 97.9), collapsing only at K=8 (92.0). So the wrist view tolerated aggressive pruning there.
LIBERO is simulation with a different camera rig and task, so SO-101 may differ, and that is exactly what this sweep measures. But do not expect top-only pruning to be safer than uniform pruning on the strength of intuition alone.
Note also the ceiling: leaving one camera intact caps the token reduction at half of
what uniform pruning reaches. With num_state_tokens=256 plus the prompt in the
sequence, this variant's 10.2% image-token cut is a smaller fraction of the
total sequence than it looks.
To attribute a result, compare against uniform pruning at a similar total token count, not only against the unpruned baseline.
Config
use_grid_token_sampler : false -> true
grid_token_sampler_num_tokens : -> [156, 0] # [top, wrist]; 0 = not pruned
action_mode : continuous -> both
setup_type : "single SO-101 5-DOF robotic arm with gripper"
The per-camera list is honoured in three places that must agree, or the image tokens scatter into the language sequence misaligned — silently, with no error: the model builds one sampler per pruned camera, the processor emits that many image placeholder tokens per camera in the same order, and the sequence-length budget sums the per-camera counts. All three were asserted equal (352) before this checkpoint was exported.
chunk_size / n_action_steps stay at 30: apply_norm_tag_metadata() pins both
from the so100_so101_molmoact2 norm tag after CLI overrides, so passing anything
else has no effect.
Training data
Merged SO-101 corpus, 177,661 frames from 11 dataset repos (5 cm teleop 21.6%, 5 cm
DAgger 11.2%, 2 cm teleop 62.1%, 2 cm DAgger 5.0%). Batches are drawn by uniform
shuffling by default; a balanced sampler pinning a 6:2:3:1 per-batch mix is opt-in
via CATEGORY_META. Normalizer statistics come from the whole corpus and are
unaffected by the sampling mode.
Status of verification
Verified at construction: config and processor agree on [156, 0], exactly one
sampler exists (for the top camera), scout_mlp.2.weight is (312, 512), the
parameter count matches the analytic prediction, and an unpruned camera passes its
grid through unchanged.
Not smoke-tested through training steps — no training was run for this variant.
Reproducing
# repo: nota-github/xpu-lerobot, branch feat/so101-train
EXPORT_INIT=1 INIT_EXPORT_DIR=/tmp/toponly-p20 GRID_NUM_TOKENS="[156,0]" HF_TOKEN=... \
scripts/train_molmoact2_so101_grid_sampler.sh
# training run it seeds
GRID_NUM_TOKENS="[156,0]" HF_TOKEN=... WANDB_API_KEY=... \
scripts/train_molmoact2_so101_grid_sampler.sh
References
- Grid Sampler (GridS) active token sampling, ICML 2026
- Fang et al. MolmoAct2: Action Reasoning Models for Real-world Deployment. arXiv:2605.02881