Open Source RoboticsRobots, software & learning

NORA

Robotics model

3.76B-parameter Vision-Language-Action model fine-tuned from Qwen2.5-VL-3B on Open X-Embodiment with the FAST action tokenizer, offering single-step (NORA) and 5-step chunking (NORA-Long) checkpoints plus RLDS and LeRobot fine-tuning scripts. [1] [2] [3] [4]

Setup & usage →CheckpointsModel hub / homepageGetting started

For
Robotics researchers exploring compact Qwen2.5-VL + FAST VLA policies; Practitioners fine-tuning 7-DoF manipulation policies on RLDS or LeRobot datasets [1] [2]
Parameters
3,758,262,272 (~3.76B BF16 parameters per HF safetensors metadata) [2] [3]
Embodiments
WidowX (BridgeData V2 real robot and OpenVLA eval server), Franka Panda (LIBERO simulation finetuned checkpoints),… [1] [2] [3]
Source rights
Named licenses are documented by part. Review licenses

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

Licenses & related hardware
Ecosystem
LeRobot

Requirements · Research details and dates

On this page

Overview

NORA (declare-lab/nora, arXiv:2504.19854) is a 3,758,262,272-parameter (~3.76B) Vision-Language-Action model developed by DeCLaRe Lab by fine-tuning Qwen2.5-VL-3B on Open X-Embodiment manipulation episodes. Rather than training a separate diffusion action head, NORA predicts discrete action tokens (vocabulary IDs 151665–153712) decoded at inference time by the physical-intelligence/fast FAST tokenizer into normalized 7-DoF end-effector deltas (x, y, z, roll, pitch, yaw, gripper). [1] [2] [4]

Model card summary

Architecture

Qwen2.5-VL-3B (Qwen2_5_VLForConditionalGeneration, 36 layers, hidden_size 2048) fine-tuned autoregressively to emit discrete robot action tokens (token IDs 151665–153712) decoded via the physical-intelligence/fast FAST tokenizer.

[2] [4]

Base model

Qwen/Qwen2.5-VL-3B-Instruct (Qwen 2.5 VL-3B)

[1] [2]

Parameter count

3,758,262,272 (~3.76B BF16 parameters per HF safetensors metadata)

Verified identical safetensors parameter count for both declare-lab/nora and declare-lab/nora-long.

[2] [3]

Input modalities
rgblanguage

Single-camera workspace RGB image and English language instruction.

[1] [2]

Output modalities
ee-pose-deltagripper-command

Normalized 7-DoF end-effector deltas (x, y, z, roll, pitch, yaw, gripper in [0, 1]).

[1] [2]

Action space

FAST-tokenized 7-DoF normalized end-effector deltas (6-DoF pose delta + gripper in [0, 1]); single-step in declare-lab/nora and 5-step action horizon in declare-lab/nora-long, unnormalized per dataset statistics.

[1] [2] [3] [4]

Trained / supported embodiments
  • WidowX (BridgeData V2 real robot and OpenVLA eval server)
  • Franka Panda (LIBERO simulation finetuned checkpoints)
  • 7-DoF EEF-delta manipulators via RLDS or LeRobot fine-tuning

[1] [2] [3]

Training datasets
  • Open X-Embodiment (OXE) RLDS robot manipulation mixture
  • LIBERO task suites (for separate LIBERO-finetuned checkpoints in the NORA HF collection)

[1] [2]

Interfaces
  • Python Nora inference class (inference/nora.py)
  • Hugging Face Transformers Qwen2_5_VLForConditionalGeneration safetensors
  • Accelerate CLI training (training/train.py, lerobot_training/lerobot_training.py)
  • RLDS and LeRobot dataset interfaces

[1] [4]

Checkpoints

CheckpointFormatSizeAccessPinned revision
declare-lab/norasafetensors7 GBopen3f373eff454d
declare-lab/nora-longsafetensors7 GBopen1fd12607a728

Two primary ungated base checkpoints (7.0 GB BF16 safetensors each); the NORA HF collection also hosts 8 LIBERO-finetuned checkpoints (4 for nora, 4 for nora-long) and successor declare-lab/nora-1.5.

[2] [3]

Is this for you?

Use cases
  • Zero-shot 7-DoF instruction following on WidowX BridgeData V2 setups
  • Fine-tuning on custom RLDS or LeRobot 7-DoF manipulation datasets with Accelerate
  • Action-chunking manipulation in LIBERO simulation via NORA-Long

[1] [2] [3]

Known limitations
  • Pretrained on a 7-DoF end-effector delta action space (6-DoF + gripper); upstream warns fine-tuning on other action spaces may not work well
  • GitHub repository declare-lab/nora has no root LICENSE file and Hugging Face model card YAML omits the license field, even though model card prose states MIT
  • NORA-Long (horizon 5) performs worse on WidowX Bridge tasks despite stronger LIBERO simulation results
  • Gripper convention is normalized to [0, 1] (0=close, 1=open), requiring the invert_grippler_action flag during training and normalize_gripper_action plus invert_gripper_action at inference for [-1, 1] environments such as LIBERO

[1] [2] [3]

Safety and limitations
  • Inference loads the external physical-intelligence/fast tokenizer with trust_remote_code=True from Hugging Face.
  • No root LICENSE file is present in declare-lab/nora on GitHub (MIT is stated only in Hugging Face model card prose), creating licensing ambiguity for formal compliance audits.
  • Pretrained exclusively on 7-DoF end-effector delta actions without proprioception or robot-side safety limits.

[1] [2] [4]

Compute & fine-tuning

Inference compute
  • Documented hardware: CUDA GPU or CPU fallback (exact GPU model and VRAM not stated in README)
  • Precision: bf16
  • Edge / consumer capable: No

7.0 GB BF16 safetensors checkpoint; exact minimum VRAM and latency_ms are not documented in README or model card.

[2] [4] · Editorial assessment

Fine-tuning
  • Supported: Yes
  • Methods: Multi-GPU RLDS fine-tuning via Accelerate (training/train.py); LeRobot dataset fine-tuning via Accelerate (lerobot_training/lerobot_training.py)
  • Recipe: Upstream fine-tuning documentation

Changing action horizon requires editing future_action_window_size in training/datasets/datasets.py.

[1]

Host computer

Single CUDA GPU (or CPU fallback in inference/nora.py) for bfloat16 inference on the 7.0 GB checkpoint; multi-GPU Accelerate cluster documented for RLDS/LeRobot pretraining and fine-tuning (exact minimum VRAM not stated upstream).

Exact GPU VRAM figures are not specified in the GitHub README or model cards.

[1] [2] [4] · Editorial assessment

Operating system
  • Linux

Standard Linux conda, CUDA, and Accelerate workflow documented.

[1] · Editorial assessment

Dependencies
  • Python 3.10
  • PyTorch / torchvision (>= 0.21.0 for LeRobot fine-tuning)
  • transformers (4.50.0 recorded in config.json)
  • accelerate (for multi-GPU training)
  • physical-intelligence/fast (external Hugging Face FAST action tokenizer loaded with trust_remote_code)
  • lerobot (optional, for LeRobot dataset fine-tuning)

[1] [4]

Skills
  • Python 3.10 and conda
  • PyTorch, Hugging Face Transformers, and Accelerate
  • RLDS or LeRobot dataset preparation

Inferred from documented conda, accelerate, RLDS, and LeRobot training workflows.

[1] · Editorial assessment

Setup & usage

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

  • Clone declare-lab/nora and create a Python 3.10 conda environment
  • Install inference/requirements.txt (or training/requirements.txt for RLDS, lerobot_training/lerobot_requirements.txt for LeRobot datasets)
  • Instantiate Nora(device='cuda') from inference.nora (downloads declare-lab/nora and physical-intelligence/fast tokenizer)
  • Call nora.inference(image, instruction, unnorm_key='bridge_orig') or pass a LeRobot Unnormalize transform

[1] [4]

Upstream-reported benchmarks

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

BenchmarkUpstream-reported scoreScope & caveats
NORA linked paper evaluations (arXiv:2504.19854; WidowX BridgeData V2 and LIBERO)No numeric benchmark tables in README or model cards (model cards note NORA performs better on WidowX Bridge and NORA-Long performs better on LIBERO; see linked arXiv:2504.19854)Qualitative comparison is upstream-reported in model cards; OSR has not inspected arXiv:2504.19854 tables or run evaluations.

[1] [3]

Project notes

Base checkpoints, horizon tradeoff, and fine-tuning routes

Two primary ungated BF16 safetensors checkpoints are hosted on Hugging Face at 7.0 GB each: declare-lab/nora (single-step prediction) and declare-lab/nora-long (pretrained with a 5-step action horizon), alongside eight LIBERO-finetuned checkpoints and a newer declare-lab/nora-1.5 release in the same collection. Upstream notes a clear empirical tradeoff: nora-long performs worse than single-step nora on WidowX Bridge tasks while performing better in LIBERO simulation. Note that the inspected inference/nora.py wrapper hard-codes the FAST tokenizer's time_horizon to 1, so users running declare-lab/nora-long should verify how five-step decoding is configured before using that wrapper for five-step predictions. Both RLDS (OpenVLA-style) and LeRobot dataset fine-tuning are supported via Hugging Face Accelerate: RLDS training runs training/train.py with accelerate launch and a cluster-matched Accelerate config, while LeRobot training runs lerobot_training/lerobot_training.py in a separate environment requiring torchvision>=0.21.0; in both routes, dataset paths and hyperparameters are configured in TrainingConfig, and changing the action horizon requires modifying future_action_window_size in training/datasets/datasets.py. Two action-space conventions matter when adapting NORA: pretraining uses 7-DoF end-effector deltas exclusively (upstream cautions that other action spaces may not fine-tune well) and pose outputs require per-robot, per-dataset unnormalization (unnorm_key or a LeRobot Unnormalize transform). In addition, NORA normalizes gripper actions to [0, 1] (0 = close, 1 = open): for environments using [-1, 1] conventions such as LIBERO (-1 = open, +1 = close), upstream documents the invert_grippler_action flag in TrainingConfig during training and the inference helpers normalize_gripper_action(..., binarize=True) followed by invert_gripper_action(...) to map predictions back to [-1, 1] with inverted sign. [1] [2] [3] [4]

Compute gaps and upstream-reported evaluations

The lightweight inference/nora.py wrapper loads declare-lab/nora in bfloat16 on CUDA (with CPU fallback) and fetches physical-intelligence/fast with trust_remote_code=True. However, neither the GitHub README nor the Hugging Face model cards report minimum GPU VRAM for inference or fine-tuning, per-step latency, or closed-loop control frequency (Hz). Likewise, the supplied README and model cards do not include numeric benchmark tables for WidowX BridgeData V2 or LIBERO, linking instead to arXiv:2504.19854 for paper results; Open Source Robotics has not inspected the paper tables or run training, fine-tuning, inference, or robot evaluations for this profile. [1] [2] [3] [4]

What we checked

None. Prepared from public upstream sources and Hugging Face API metadata without hands-on training, fine-tuning, or inference.

Not yet established

  • Documented closed-loop control frequency (Hz)

Updates

Upstream last updated
Jan 9, 2026 [1]
Latest release
Aug 27, 2025 [2] [3]
Profile updated
Nov 21, 2026
Sources last checked
Nov 21, 2026
Next review
Feb 19, 2027
Reviewed revision
declare-lab/nora@6b18c23d7875052e03fba4f8c2f32bd6a8a5c4a9; HF declare-lab/nora@3f373eff454d299bad4aa0d845a5c0db4fb0e0bc · combined hardware and software repository

Separate license matrix

Open-weights 3.76B Qwen2.5-VL + FAST VLA with ungated safetensors checkpoints and RLDS/LeRobot training code; classified as partial openness because declare-lab/nora lacks a root LICENSE file and HF YAML metadata omits a machine-readable license tag (MIT is stated only in model card prose).

LayerAvailability / license
Codeavailable · MIT (declared in HF model card prose; no root LICENSE file in GitHub repo)
GitHub API reports license: null and root directory contains no LICENSE file; HF model cards explicitly state 'All Nora checkpoints, as well as our training codebase are released under an MIT License.' [1] [2]
Model weightsavailable · MIT (declared in HF model card prose; omitted from HF YAML frontmatter)
HF API cardData.license is null because YAML frontmatter omits license:, though model card body lists License: MIT. [2] [3]
Training datasetsavailable · Mixed (Open X-Embodiment RLDS mixture; per-dataset upstream licenses apply)
[1] [2]
Documentationavailable · MIT (declared in HF model card prose; no root LICENSE file in GitHub repo)
[1] [2]

Licensing hygiene gap between GitHub and Hugging Face

A notable compliance caveat separates NORA's prose claims from its repository metadata: both the declare-lab/nora and declare-lab/nora-long model card bodies state that all NORA checkpoints and the training codebase are released under an MIT License, yet the declare-lab/nora GitHub repository has no root LICENSE file (license: null via GitHub API) and the Hugging Face YAML frontmatter omits a license: tag. Teams with strict license-scanning pipelines should verify license intent directly with the maintainers and audit individual Open X-Embodiment dataset licenses before commercial deployment. [1] [2] [3]

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 — Initial 0.8.0-draft.1 model profile prepared from declare-lab/nora@6b18c23d7875 and HF declare-lab/nora + nora-long checkpoints.