Open Source RoboticsRobots, software & learning

MolmoBot

Robotics model

Robot manipulation policies and evaluation code paired with simulated training data and the MolmoSpaces generation engine. [1] [3]

Setup & usage →CheckpointsModel hub / homepageSource codeGetting started

For
Robot learning researchers comparing simulation-trained manipulation policies. [1]
Parameters
~4B VLM backbone (Molmo2-4B) + per-layer action expert (18.62 GB PyTorch checkpoint for MolmoBot-DROID [6] [8]
Embodiments
DROID (Franka Emika Panda via Polymetis), Rainbow Robotics RBY1 (dedicated RBY1 checkpoints), MolmoSpaces / MuJoCo… [1] [3] [9]
Source rights
Named licenses are documented by part. Review licenses

Source researched · Not tested by us
Sources checked · Research details

Licenses & related hardware
Ecosystem
MuJoCo

Requirements · Research details and dates

On this page

Overview

MolmoBot (allenai/MolmoBot) is a family of manipulation policies from the Allen Institute for AI trained on procedurally generated simulation episodes (allenai/MolmoBot-Data via MolmoSpaces) without real-robot training data for its primary DROID checkpoint. At the pinned Hugging Face revision (cbe6ec358958), allenai/MolmoBot-DROID (18.62 GB PyTorch model.pt on disk) builds on allenai/Molmo2-4B (Qwen3-4B-Instruct + SigLIP 378px) with a per-layer cross-attention flow-matching action expert (action_dim: 8, action_horizon: 16, n_action_steps: 8, flow_matching_num_steps: 10, quantile normalization). [1] [6] [8]

Model card summary

Architecture

Flow-matching VLA initialized from Molmo2-4B (Qwen3-4B-Instruct LLM + SigLIP 378px vision encoder) with a per-layer cross-attention action expert (10 flow-matching steps); companion variants include MolmoBot-Pi0 (PaliGemma-3B) and MolmoBot-SPOC.

[1] [6] [8] [9]

Base model

allenai/Molmo2-4B (MolmoBot-DROID / RBY1); google/paligemma-3b-mix-224 (MolmoBot-Pi0-DROID)

[6] [8] [9]

Parameter count

~4B VLM backbone (Molmo2-4B) + per-layer action expert (18.62 GB PyTorch checkpoint for MolmoBot-DROID; exact safetensors count not published)

Inferred from Molmo2-4B base model and 18.62 GB model.pt; binaries not downloaded.

[6] [8] · Editorial assessment

Input modalities
rgbproprioceptionlanguage

Wrist camera + 1 exo RGB camera, robot joint state (observation.state), and natural-language task instruction.

[1] [8]

Output modalities
joint-positionsgripper-command

Joint-position + gripper actions (--action-type joint_pos, action_dim: 8 for DROID Franka).

[1] [8]

Action space

Continuous flow-matching action chunks (action_dim: 8, action_horizon: 16, n_action_steps: 8, flow_matching_num_steps: 10, quantile normalization) for DROID joint-position and gripper control.

[1] [8]

Trained / supported embodiments
  • DROID (Franka Emika Panda via Polymetis)
  • Rainbow Robotics RBY1 (dedicated RBY1 checkpoints)
  • MolmoSpaces / MuJoCo simulation

[1] [3] [9]

Training datasets
  • allenai/MolmoBot-Data (procedurally generated simulation episodes via MolmoSpaces; zero real-robot data for MolmoBot-DROID)

[6] [7]

Interfaces
  • Policy server CLI (launch_scripts/serve_molmo.py --hf-repo ... --action-type joint_pos)
  • DROID / Polymetis robot evaluation client
  • MolmoSpaces / MuJoCo simulation eval

[1] [3]

Checkpoints

CheckpointFormatSizeAccessPinned revision
allenai/MolmoBot-DROIDpytorch18.62 GBopencbe6ec358958
allenai/MolmoBot-RBY1Multitaskpytorch18.62 GBopen0786434b1014
allenai/MolmoBot-Pi0-DROIDsafetensors6.53 GBopen48c8e2fc1868
allenai/MolmoBot-SPOC-DROIDsafetensors0.41 GBopen69b48b801478

Four representative ungated checkpoints verified via Hugging Face API; note variant licenses differ (Apache-2.0 vs Gemma vs Other).

[8] [9]

Is this for you?

Use cases
  • Simulation policy evaluation
  • DROID manipulation policy research

[1]

Known limitations
  • Paper performance is author-reported and not reproduced here.
  • DROID evaluation needs matching hardware, cameras, Polymetis servers and checkpoint.
  • Dataset scenes carry object-specific reuse considerations.

[1] [3] [7]

Safety and limitations
  • DROID deployment requires matching Franka/Polymetis robot and gripper servers, wrist + exo cameras, and --action-type joint_pos.
  • Companion checkpoints carry different licenses (Apache-2.0 for MolmoBot-DROID/RBY1, Gemma for MolmoBot-Pi0-DROID, Other for MolmoBot-SPOC-DROID).
  • MolmoBot-Data carries an ODC-BY dataset card, while individual simulated 3D scene objects have asset-specific licenses; commercial-use selection should consult the provided non-NC episode filter and applicable object terms.

[1] [6] [7] [9]

Compute & fine-tuning

Inference compute
  • Documented hardware: CUDA GPU (depends on variant; 18.62 GB model.pt for MolmoBot-DROID)
  • Edge / consumer capable: No

Exact minimum inference VRAM and latency_ms not stated upstream.

[1] [8]

Fine-tuning
  • Supported: Yes
  • Methods: Full training / fine-tuning from Molmo2-4B via torchrun (8-64 H100s in upstream recipe; micro-batching supported)
  • Recipe: Upstream fine-tuning documentation

[3]

Host computer

GPU resources depend on policy variant and task; no independently measured minimum.

[1]

Operating system
  • Linux

uv / MuJoCo EGL / Polymetis Linux workflow documented.

[3] · Editorial assessment

Dependencies
  • MolmoSpaces for data generation
  • DROID/Polymetis for documented physical evaluation

[1]

Skills
  • Python
  • PyTorch
  • uv
  • Polymetis / DROID or MuJoCo

Inferred from documented uv, MuJoCo simulation, and Polymetis/DROID evaluation instructions.

[1] [3] · Editorial assessment

Setup & usage

Open the upstream getting-started guide [1] · Model card / documentation [1]

  • Follow the policy README for the simulation demo.
  • For DROID, configure the matching camera and Polymetis services before starting policy server and evaluation client.

[1] [3]

Upstream-reported benchmarks

Upstream-reported only: Open Source Robotics has not run training, inference, or hardware evaluations for these benchmarks.

BenchmarkUpstream-reported scoreScope & caveats
Zero-shot sim-to-real manipulation on DROID and RBY1 (arXiv:2603.16861)See arXiv:2603.16861 (upstream reports zero-shot sim-to-real transfer trained purely on MolmoBot-Data simulation episodes)Author-reported in arXiv:2603.16861; not independently verified by OSR.

[1] [6]

Project notes

Checkpoints, workflows, and companion variants

Upstream publishes multiple ungated checkpoints across Franka DROID and Rainbow Robotics RBY1 embodiments, including allenai/MolmoBot-DROID (18.62 GB), allenai/MolmoBot-RBY1Multitask (18.62 GB), allenai/MolmoBot-Pi0-DROID (6.53 GB safetensors, PaliGemma-3B base), and allenai/MolmoBot-SPOC-DROID (0.41 GB safetensors). For software inspection, MolmoBot/README.md documents uv sync --extra eval, a simulation demo notebook, and MuJoCo/MolmoSpaces evaluation. Physical DROID execution requires Polymetis robot and gripper servers, one wrist and one external RGB camera, a policy server (serve_molmo.py --hf-repo allenai/MolmoBot-DROID --action-type joint_pos), and the robot evaluation client. Minimum inference VRAM, inference latency, and deployment control frequency are not established by this review, and checkpoint download sizes are disk sizes rather than GPU-memory requirements. [1] [3] [8] [9]

Training compute and upstream-reported benchmarks

Upstream's replication recipe trains from Molmo2-4B using torchrun on 8–64 H100 GPUs (noting micro-batching support for smaller GPUs). Zero-shot sim-to-real transfer results on DROID and RBY1 in arXiv:2603.16861 are author-reported; Open Source Robotics has not downloaded checkpoint binaries or data shards, and has not run training, fine-tuning, inference, simulation, or robot evaluations. [1] [3] [6]

What we checked

Source research only. No commercial relationship or hands-on testing. No upstream media copied to the public site.

Not yet established

  • Documented closed-loop control frequency (Hz)

Updates

Upstream last updated
Apr 27, 2026 [1]
Latest release
Not established
Profile updated
Nov 21, 2026
Sources last checked
Nov 21, 2026
Next review
Feb 19, 2027
Reviewed revision
Commit snapshot 33c0ca77bf6062a23d60ffd4a6859334c4a46d30; no matching GitHub release found · combined hardware and software repository

Separate license matrix

Inspectable robot-specific policy implementation with simulation and DROID workflows; data-generation engine, weights and dataset retain separate scope and conditions.

LayerAvailability / license
Codeavailable · Apache-2.0
Artifact and declaration inspected at the cited scope. External assets/dependencies and actual operation need separate verification. [3] [2]
Model weightsavailable · Apache-2.0 (MolmoBot-DROID / RBY1); Gemma (MolmoBot-Pi0-DROID); Other (MolmoBot-SPOC-DROID)
Primary Molmo2-4B-based checkpoints declare Apache-2.0, while companion Pi0 and SPOC variants carry Gemma and Other license tags on Hugging Face Hub. [6] [8] [9]
Training datasetsavailable · ODC-BY card declaration; object-level restrictions require filtering
Artifact and declaration inspected at the cited scope. External assets/dependencies and actual operation need separate verification. [7]
Documentationavailable · Apache-2.0
Artifact and declaration inspected at the cited scope. External assets/dependencies and actual operation need separate verification. [1] [2]

Separate code, variant weight, and dataset rights

The allenai/MolmoBot code repository is Apache-2.0, and the primary Molmo2-4B checkpoints (MolmoBot-DROID, MolmoBot-RBY1Multitask) declare Apache-2.0. Companion checkpoints carry different Hugging Face license tags (gemma for MolmoBot-Pi0-DROID, other for MolmoBot-SPOC-DROID). allenai/MolmoBot-Data declares ODC-BY at the dataset card level, while individual simulated 3D scene objects carry asset-specific licenses; users selecting episodes for commercial use should consult the provided non-NC episode filter and applicable object licenses rather than treating the dataset label alone as blanket clearance. [2] [6] [7] [9]

References

Each source links to the exact revision we read. “Used for” shows where this page relies on it.

Correction guidance · Discussion is not enabled · Profile format 0.8.0-draft.1

Change history

2026-11-21 — Migrated 0.7.0-draft.5 -> 0.8.0-draft.1 (kind=model); added verified config.yaml (S008) and companion checkpoint metadata (S009) while preserving historical source dates.