NORA-1.5
SUTD · November 2025
DeCLaRe Lab
- Parameters
- 4 B
- The Hugging Face model card lists 4B parameters in F32.
- GPU memory
- 7.5 GB
- weights at bf16, computed. Fits a 12 GB card
- Inference latency
- not published
- per action step
- Weights
- MIT
- downloadable checkpoint
What it is
NORA-1.5 keeps the 3B Qwen2.5-VL backbone of NORA and adds a flow matching action expert plus reward driven post-training, reaching roughly 4B parameters in total. The post-training uses direct preference optimization with two reward signals, an action conditioned world model that scores goal progress and a heuristic measuring deviation from ground truth actions. It lifts the LIBERO average from 79.5 to 94.5 percent (95.0 with DPO) and raises SimplerEnv Google Robot visual matching from 76.9 percent zero-shot to 82.8 percent, and on a real Galaxea A1 arm the authors report 83.84 percent against 44.44 percent for pi0. The weak spot is the visually shifted SimplerEnv split at 71.9 percent, about 11 points below the clean split. For a deployment decision the gap in the record matters: no latency, control frequency or VRAM figure is published anywhere, so the NORA figure of 8.3 GB cannot be reused for this model.
Architecture
- Backbone
- NORA (Qwen2.5-VL-3B), initialised from declare-lab/nora-long
- Action head
- flow matching action expert reading from the NORA backbone and generating action sequences directly. A FAST+ tokenised variant, NORA-1.5-FAST, is evaluated alongside it. Post-training uses direct preference optimization driven by an action conditioned world model reward plus a deviation from ground truth heuristic.
- Action chunk
- 5 steps per forward pass
- Parameters
- The Hugging Face model card lists 4B parameters in F32. That is the 3B NORA backbone plus the added flow matching action expert. The paper does not state a single combined parameter figure.
- Pretraining data
- Open X-Embodiment, inherited from the NORA backbone. Real robot fine-tuning in the paper used 50K frames on a Galaxea A1 and 1,000 teleoperated pick and place episodes with randomised object placement.
- Embodiments
- Google Robot (SimplerEnv), WidowX / BridgeData V2, LIBERO (simulation), Galaxea A1 (real)
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.
| Precision | Weights | Note |
|---|---|---|
| fp32 | 15 GB | Training master weights, rarely used for inference |
| bf16 or fp16 | 7.5 GB | The usual way these checkpoints are served |
| int8 | 3.7 GB | Quantised, expect some loss of precision on fine motions |
| int4 | 1.9 GB | Aggressive quantisation, verify success rate before trusting it |
Smallest real card that fits the bf16 weights plus a 40 percent runtime allowance: 12 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.
No VRAM, latency or control frequency figure is published for NORA-1.5. The 8.3 GB inference figure from the original NORA paper covers the 3B backbone only and does not include the added flow matching action expert, so it must not be carried over.
The authors publish no memory or latency figure for this model. Everything above is computed from the parameter count.
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-1.5 with DPO post-training95%success rateAuthors' own evaluation. Per suite: spatial 98.0, object 96.0, goal 95.4, long 90.5.source
- LIBERO average over spatial, object, goal and long, NORA-1.594.5%success rateAuthors' own evaluation. Per suite: spatial 97.3, object 96.4, goal 94.5, long 89.6. NORA scored 79.5 in the same table.source
- SimplerEnv Google Robot, visual matching, with DPO post-training82.8%
- SimplerEnv Google Robot, visual matching, zero-shot76.9%success rateAuthors' own evaluation. Per task: pick coke can 85.6, move near 88.6, open/close drawer 56.7.source
- SimplerEnv Google Robot, variant aggregation, with DPO post-training71.9%success rateAuthors' own evaluation. The harder, visually shifted split. 10.9 points below the visual matching score, which is where robustness to appearance change shows up.source
Real world
No results in this category are published for this model.
Fine tuned tasks
No results in this category are published for this model.
On real hardware
Evaluated on a Galaxea A1 arm across nine tasks after fine-tuning on 50K frames. The authors report NORA-1.5-FAST with DPO post-training at 83.84 percent success, NORA-1.5-FAST at 78.88 percent, NORA at 73.3 percent and pi0 (3.3B) at 44.44 percent, an improvement of 13 to 46 points over the baselines. This is one arm in one lab, so it should be read as evidence that the post-training helps, not as a cross platform guarantee.
Fine tuning it yourself
Inference runs through the VLAWithExpert class, calling model.sample_actions(image, instruction, num_steps=10). Real robot fine-tuning in the paper used 50K frames on a Galaxea A1 plus 1,000 teleoperated pick and place episodes. Reproducing the reward driven post-training additionally requires an action conditioned world model to score candidate actions, which is a substantial extra component beyond ordinary imitation fine-tuning.
Where it helps, where it does not
Strengths
- LIBERO average of 94.5 percent (95.0 with DPO), which puts a 3B class open model level with pi0 at 94.2.
- LIBERO-Long at 89.6 percent, the long horizon suite where most compact policies collapse.
- Flow matching action expert replaces NORA's single token autoregressive decoding, so it emits chunks of 5 actions.
- Consistent gains over NORA on every reported benchmark, plus 13 to 46 point gains over baselines on a real Galaxea A1.
- MIT licensed with code and weights on the Hub.
Limits
- No published latency, control frequency or memory figure, even though inference speed is one of the stated motivations for the flow matching expert.
- Reproducing the headline gains requires DPO post-training against a learned world model reward, which is extra machinery most teams will not rebuild.
- SimplerEnv variant aggregation drops to 71.9 percent against 82.8 percent on visual matching, so robustness to visual change remains the weak axis.
- Real robot evidence is a single Galaxea A1 arm from the authors' own lab.
- At publication the model card noted that code and full checkpoints were still pending release, so availability should be re-checked before committing.
- Only three months of community use at the time of writing, so there is little independent reproduction.
Sources
Everything on this page was taken from these documents. Where they disagree with what you read here, they win.
- https://arxiv.org/abs/2511.14659
- https://arxiv.org/html/2511.14659
- https://huggingface.co/declare-lab/nora-1.5
- https://github.com/declare-lab/nora-1.5
- https://declare-lab.github.io/nora-1.5
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.