What Is NORA?

NORA is introduced in NORA: A Small Open-Sourced Generalist Vision Language Action Model for Embodied Tasks (arXiv 2504.19854), posted on 28 April 2025 by eight authors from Singapore University of Technology and Design and Lambda Labs. The abstract argues that existing vision-language-action (VLA) models can fail at tasks such as object grasping because of limits in visual encoding, and that they are often larger than 7B parameters, which makes them impractical for real-time robotic environments. The PDF and project page add that VLA models are commonly fine-tuned for specific tasks, so there is a need for a smaller model that can be fine-tuned on consumer-grade GPUs. NORA, a 3B-parameter model, is the authors’ answer: lower computational overhead while keeping strong task performance. This page draws only on the paper, the project page, the official GitHub repository and the Hugging Face model card. NORA is a robot-control policy trained on robot demonstrations, so egocentric human video as training data is Not stated in the sources read.

Per the paper, NORA builds on a pretrained vision-language model, Qwen-2.5-VL-3B, chosen for its balance of performance and efficiency at that scale. The input is a natural language task instruction plus a visual observation, which the paper keeps to a single frame, resized to 224 by 224 pixels for training. The model predicts an action chunk autoregressively as discrete tokens, and the FAST+ decoder turns them back into continuous actions. The VLM tokenizer’s vocabulary is augmented by 2048 tokens from FAST+, and training uses a standard next-token prediction loss. NORA uses an action chunk size of 1. A variant, NORA-Long, shares the architecture but predicts five actions, and was pretrained for 900k steps on the same data. The paper states that inference uses about 8.3 GB of GPU memory. The Hugging Face card describes the output as normalized 7-DoF end-effector deltas (x, y, z, roll, pitch, yaw, gripper), un-normalized per robot and dataset, and lists the model size as 4B parameters against the paper’s 3B; the sources read do not reconcile the two.

A language backbone needs actions as tokens. The paper describes the common binning approach, where each action dimension maps to one of 256 bins, and notes that OpenVLA overwrites the 256 least-used tokens in its language tokenizer. NORA instead adopts the FAST tokenizer, which applies a discrete cosine transform across the action dimensions at each timestep to decorrelate them, then compresses the result with byte-pair encoding into shorter sequences. The paper says this reduces vocabulary size and accelerates convergence, and it picked FAST+ for its stated efficacy across single-arm, bi-manual and mobile robot tasks. The FAST action tokenizer entry covers the tokenizer itself, so this page does not restate it.

The GitHub README adds practical limits: NORA is pretrained on a 7-DoF action space (six plus a gripper), fine-tuning on other action spaces may not work well, and gripper actions are flipped and normalized to 0 (close) to 1 (open) in pretraining. The README also says training supports the LeRobot dataset format.

The abstract says NORA is trained on 970k real-world robot demonstrations. The body says it is pretrained on the Open X-Embodiment (OXE) dataset, trajectories from different robots, including subsets such as BridgeV2 and DROID. The pre-training mixture split is in the paper’s appendix and is not reproduced here. Training took roughly three weeks on a single node of eight H100 GPUs, about 4000 H100 GPU hours, with batch size 256, 1.1 million gradient updates and AdamW.

For evaluation the authors use a real-world WidowX robot and the LIBERO simulation benchmark. Over nine real-world tasks, each run for 10 trials, they report an average success rate of 56.7% for NORA against 40 for OpenVLA, 11.1 for SpatialVLA and 4.4 for RT-1, and 87.9% on LIBERO for fine-tuned NORA-Long. These are the authors’ own results on their own setups, and Firsthand draws no comparison from them. They also report that NORA and OpenVLA both degrade when distractors are added, and that executing all five NORA-Long actions on the WidowX made the robot crash into its environment. Egocentric human video is not described as part of the training data, and dataset licences are Not stated in the sources read. The model card states MIT for the checkpoints and training codebase.

The NORA sources read do not compare it with SmolVLA or TinyVLA, so the first comparison table sets side by side only what each official source or existing Firsthand entry states. Firsthand ranks none of them. The structural differences are plain: NORA starts from a 3B vision-language model and emits discrete FAST+ tokens, SmolVLA adds a flow matching action expert to a smaller model and outputs continuous actions, and TinyVLA adds a diffusion policy decoder during fine-tuning and says it skips robot pretraining. Pretraining data differs too, with Open X-Embodiment trajectories for NORA and 481 community datasets for SmolVLA. Egocentric human video is not described for any of the three. Read the SmolVLA and TinyVLA entries for their own sources.

NORA is a public research release, not a service. Firsthand is not affiliated with its authors and makes no claim that its data trained, was used with or is compatible with NORA. NORA’s sources describe robot demonstrations with 7-DoF end-effector actions, while Firsthand captures human first-person video, not robot data, so robot joint state and robot actions are Not stated. If you need your own tasks and environments, the site states custom egocentric collection to a written spec, fast and on demand: 3D hand pose at 21 joints per hand with a per-joint visibility flag, wrist trajectories and action segments, from a head-mounted camera plus two wrist cameras on one hardware-triggered clock, alongside depth and IMU. The only language field is one task instruction per episode.

Confirm a spec on a Spec Pilot, a small fast first batch. The median from signed spec to first validated batch is 48 hours where coverage is deep and three to four weeks where it is thin. Billing is per validated hour; delivery is RLDS, LeRobot v2.1, WebDataset or HDF5. Mapping hand pose to NORA’s action space would be the buyer’s pipeline, and this page does not claim it is possible. If contact signals matter, ask for the separate pilot-ready tactile glove tier.

What is NORA?

A published 3B-parameter, open-source vision-language-action model built on Qwen-2.5-VL-3B, which uses the FAST+ tokenizer to generate robot actions, per arXiv 2504.19854.

How many parameters does NORA have?

The paper states 3B parameters. The Hugging Face model card lists 4B params, and the sources read do not reconcile the two.

What data was NORA trained on?

The abstract says 970k real-world robot demonstrations, from the Open X-Embodiment dataset. Egocentric human video is not described as part of the training data.

How does NORA represent actions?

As discrete tokens from the FAST+ tokenizer, which uses a discrete cosine transform and byte-pair encoding, decoded back to continuous actions. The model card describes 7-DoF end-effector deltas.

Is NORA open source?

The model card states an MIT licence for the checkpoints and training codebase. Licences of the training datasets are Not stated in the sources read.

Does Firsthand use NORA, and can I request a custom dataset?

Firsthand is not affiliated with the NORA authors and makes no claim of use, training or compatibility. You can request a custom egocentric dataset: write a spec, confirm it on a Spec Pilot, pay per validated hour.