ccwatson/MolmoBot-DROID-Directional-PickPlace-SimpleDesk-Step-10000

MolmoBot-DROID Directional Pick/Place SimpleDesk — Step 10,000

This repository contains the 10,000-step intermediate checkpoint from the lr_custom_16k MolmoBot finetuning run. It was finetuned from allenai/MolmoBot-DROID for Franka/FR3 tabletop pick-and-place commands that require placing one object to the left or right of another object.

The root config.yaml and model.pt form the unsharded MolmoBot checkpoint expected by the MolmoBot server. This is not the final 16,000-step checkpoint; the step is deliberately included in both the repository name and this model card.

Checkpoint identity

Field	Value
Checkpoint step	10,000
Training run	lr_custom_16k
Checkpoint format	MolmoBot unsharded (config.yaml + model.pt)
Base model	allenai/MolmoBot-DROID
Base model revision	cbe6ec358958d07ddfb20d3aa54e560e9e1b18c9
Training W&B run	ccwatson/molmobot_finetune/ouuzdjlg
Training Slurm job	7107997 (lr-custom-ie-fe-16k-b16-32)
Checkpoint created	2026-07-17 12:11 UTC
model.pt size	19,992,166,548 bytes
model.pt SHA-256	027daf40cbf87cfebc4dcee0aa16b0d6c0a87c2bfb30229ed77f9c1545d600da
config.yaml SHA-256	7eeb51134b93fc17955f3bd850240074d6b9faf3dcf527177861940c7c098400

The complete resolved model/trainer configuration is in config.yaml. The submission-time environment is preserved as metadata/training_launch.env, and machine-readable provenance is in metadata/provenance.json.

Training data

Training used ccwatson/directional_pick_place_simpledesk_diverse_targets_tested40_200demos_20260716_molmobot at dataset revision 5446191605f7eb3020983d83f707153d9719ecfb.

The dataset contains 200 successful scripted trajectories derived from 40 valid benchmark episodes, with five attempts per episode:

Split	Source episodes	Trajectories	HDF5 files	Videos
Train	36	180	180	360
Validation	4	20	20	40

The two cameras are droid_shoulder_light_randomization and wrist_camera_zed_mini. Prompts were not randomized during training. Normalization was computed from the train split using action quantiles and state min/max statistics for the eight-dimensional Franka action and state vectors. The exact statistics are included as normalization_stats.yaml (SHA-256 972a88e10f02976ce5a4d1c90a2070a62e4daa08a5fe1cecb4ca00e8216eb15d).

Main training settings

Setting	Value
Intended run length	16,000 steps
This checkpoint	10,000 steps
Device / global batch size	16 / 32
Sequence length	928
Observation frames	2
Observation step delta	8
Maximum images	4
Maximum frames	1
Single-frame preprocessing	false
Action preset	franka_joint
Camera preset	franka_droid
Action horizon in data	16 × 8
Weighted sampling	enabled
Prompt randomization	disabled
Image augmentation	enabled
LLM finetuning	enabled
Embedding finetuning mode	ae
Flow timestamps	8
Connector learning rate	5e-6
Vision encoder learning rate	5e-6
LLM learning rate	1e-5
Action expert learning rate	1e-4
Validation interval / max examples	200 / 200
Save and external-eval interval	2,000 steps
Trainer and optimizer state	reset from base model
Hardware	1× NVIDIA B200, 28 CPUs, 224 GiB host memory

Closed-loop evaluation

This checkpoint was evaluated on all 40 episodes of FrankaPickAndPlaceRelativeNextToSimpleDeskDiverseTargetsTested40Bench (20 left and 20 right tasks) through the VLA evaluation harness and MolmoSpaces. The evaluation used natural-language prompts, Filament rendering, 10 parallel shards, a 500-step task horizon, and both DROID cameras.

Server action settings were:

• execution horizon: 8
• state mode: cross_attn
• maximum joint delta: 0.2

Metric	Result
Strict success	25 / 40 (62.5%)
Lax success	26 / 40 (65.0%)
Strict oracle-done rate	25 / 40 (62.5%)
Lax oracle-done rate	27 / 40 (67.5%)
Average episode length	274.675 steps
Average successful episode length	140.08 steps
Min / max episode length	32 / 499 steps
Failed shards	0
Missing / duplicate / unexpected episodes	0 / 0 / 0

The eval completed successfully (eval_returncode=0) in 2,016.92 seconds. See the evaluation W&B run, metadata/external_eval_request.json, and metadata/external_eval_result.json for the complete recorded configuration and result.

Repository files

• model.pt: unsharded MolmoBot weights.
• config.yaml: complete resolved model and trainer configuration.
• normalization_stats.yaml: exact state/action normalization statistics.
• metadata/training_launch.env: exact experiment launch configuration.
• metadata/external_eval_request.json: checkpoint identity and requested eval settings.
• metadata/external_eval_result.json: complete eval result and coverage accounting.
• metadata/provenance.json: source revisions, hashes, run identity, and artifact inventory.

Intended use and limitations

This checkpoint is intended for research evaluation of the matching Franka/FR3 DROID-camera setup and directional SimpleDesk tasks. It should be served with the MolmoBot inference stack and the included normalization stats. It has only been measured on the 40-episode simulated benchmark described above. The 62.5% result is a single benchmark evaluation, not a claim of broad robotic capability or real-world safety. Validate action scaling, cameras, workspace limits, collision handling, and emergency-stop behavior before any physical-robot use.

License

This derivative checkpoint follows the Apache-2.0 license declared by the base allenai/MolmoBot-DROID model. Users remain responsible for the terms of the base model, training dataset, simulator assets, and other dependencies.