SUT

NORA

SUTD · April 2025

DeCLaRe Lab

Compact VLAOpen weights, MITAlso written NORA-Long, declare-lab/nora
Parameters
3 B
3B parameters, stated in the abstract.
GPU memory
8.3 GB
reported by the authors
Inference latency
not published
per action step
Weights
MIT
MIT.

What it is

NORA is a 3B parameter generalist vision language action model built by fine-tuning Qwen2.5-VL-3B on 970k Open X-Embodiment demonstrations, using the FAST+ tokenizer to emit discrete action tokens. Its stated target is the gap between 7B class policies that are too slow for real time control and small policies that lack world knowledge. On the authors' own WidowX arm it averaged 56.7 percent over nine tasks against 40 percent for OpenVLA, with its biggest margin on out of distribution object grasping. The base model predicts a single action per forward pass, and the NORA-Long chunking variant that fixes this in simulation (87.9 percent LIBERO average) actually performed worse on the real arm. It is fully MIT licensed and reports about 8.3 GB of inference memory, but publishes no latency or control frequency, and it has since been superseded by NORA-1.5.

Architecture

Backbone
Qwen2.5-VL-3B
Action head
autoregressive discrete action tokens produced by the FAST+ tokenizer, which is DCT based
Action chunk
1 step per forward pass
Parameters
3B parameters, stated in the abstract. The FAST+ action tokenizer adds 2048 tokens to the vocabulary. NORA-Long shares the same 3B architecture but predicts chunks of 5 actions and was pretrained for 900k steps instead of 1.1M.
Pretraining data
970k real world robot demonstrations from Open X-Embodiment, including BridgeV2 and DROID subsets, at 224x224 resolution. Training ran on a single node of 8x H100 for roughly three weeks, about 4,000 H100 GPU hours, at batch size 256 for 1.1 million gradient updates with AdamW, peak learning rate 5e-5, bf16 and FlashAttention.
Embodiments
WidowX (BridgeData V2, real), LIBERO (simulation), Open X-Embodiment mixtures

What hardware it needs

This is the section most people came for, so it is worth being precise about which numbers are measured and which are arithmetic.

Computed from 3 B parameters. Weights only, so treat it as a floor: activations, image encoder intermediates and the CUDA context come on top, in practice 30 to 50 percent more for inference.
PrecisionWeightsNote
fp3211 GBTraining master weights, rarely used for inference
bf16 or fp165.6 GBThe usual way these checkpoints are served
int82.8 GBQuantised, expect some loss of precision on fine motions
int41.4 GBAggressive quantisation, verify success rate before trusting it

Smallest real card that fits the bf16 weights plus a 40 percent runtime allowance: 8 GB. That is an estimate for inference, not for fine tuning. Full fine tuning also holds optimiser state and gradients, which typically costs several times the weights unless you use LoRA or a comparable adapter.

Published figures. These come from the model authors, not from this site.
Inference memory, reported8.3 GB

The paper reports about 8.3 GB of GPU memory at inference, which fits a 12 GB consumer card, but names no GPU for that measurement. Pretraining used 8x H100 80GB for roughly three weeks. The stated design goal is fine-tuning on consumer grade GPUs, but no fine-tuning VRAM figure is given.

Reported results

Grouped by what the evaluation actually asked. Values from different suites are not comparable with each other, so each one keeps its suite and split.

Simulation

Reproducible benchmark suites. Everyone runs the same episodes, so the numbers can be compared inside a suite.

  • LIBERO average over spatial, object, goal and long, NORA-Long fine-tuned
    87.9%
    success rateAuthors' own evaluation, 500 trials per suite. Per suite: spatial 92.2, object 95.4, goal 89.4, long 74.6.source
  • LIBERO average over spatial, object, goal and long, base NORA fine-tuned
    73.9%
    success rateAuthors' own evaluation. Below OpenVLA fine-tuned at 76.5. The action chunking variant NORA-fine-tuned-AC reaches 79.5.source
  • LIBERO long, base NORA fine-tuned
    45%
    success rateAuthors' own evaluation. Weakest suite for the base model, well below OpenVLA at 53.7.source

Real world

Rollouts on physical hardware. The setups differ, so read these as evidence, not as a ranking.

  • Real-world WidowX (authors' setup) 9 tasks, 10 trials each, overall average
    56.7%
    success rateAuthors' own hardware. Same table reports OpenVLA at 40, SpatialVLA at 11.1 and RT-1 at 4.4. NORA's out of distribution object group averaged 83 percent.source

Fine tuned tasks

No results in this category are published for this model.

On real hardware

Authors' own WidowX setup, nine tasks with ten trials each. NORA averaged 56.7 percent overall against 40 percent for OpenVLA, 11.1 percent for SpatialVLA and 4.4 percent for RT-1. Its strongest group was out of distribution object grasping at 83 percent average, and its weakest was scenes with multiple objects at 33 percent. The NORA-Long chunking variant was better in simulation but worse on the real arm, because its affordance point estimation degraded on smaller objects.

Fine tuning it yourself

Fine-tuned on LIBERO format demonstrations. The authors publish per suite fine-tuned checkpoints on the Hub, for example declare-lab/nora-finetuned-libero-object and declare-lab/nora-long-finetuned-libero-goal, so the released artefacts show the intended per task fine-tune pattern. Neither paper nor model card states a minimum demonstration count.

Where it helps, where it does not

Strengths

  • 3B parameters at MIT license with full training code released, which is unusually open for a generalist policy.
  • Beats OpenVLA (7B) on a real WidowX arm, 56.7 against 40 percent over nine tasks.
  • Strongest on out of distribution object grasping at 83 percent, which is the failure mode the paper set out to fix.
  • Documented inference memory of about 8.3 GB, one of the few models in this group with any published VRAM number.
  • NORA-Long reaches 87.9 percent on the LIBERO average, competitive with much larger policies.

Limits

  • Base NORA predicts one action per forward pass, which caps achievable control rate.
  • NORA-Long improves simulation results but got worse on the real WidowX arm, so the chunking gain does not transfer across embodiments.
  • Base model LIBERO-Long at 45 percent is weak for long horizon tasks and trails OpenVLA.
  • No inference latency and no control frequency is published, only a memory figure without a named GPU.
  • Superseded by NORA-1.5 in November 2025, which reports 94.5 percent on the LIBERO average against 79.5 for the NORA chunking variant.
  • Real robot evidence is one WidowX arm from a single lab.

Sources

Everything on this page was taken from these documents. Where they disagree with what you read here, they win.

Entry last checked 2026-08-11.

Try a policy on a real arm

A physical SO-100 is online and free to drive in the browser, and five of the models in this arena can be fine tuned on a dataset you record with your own.