VLA Gallery / All models / Flow·Diffusion VLA
NORA-1.5
Nanyang Technological University, Lambda Labs, SUTD · 2025-11
VLAFlow·Diffusion VLATotal not disclosedhybrid
Attaches a flow expert to an autoregressive VLA, jointly learning FAST+ tokens and continuous actions, plus DPO reward post-training.
Paper
NORA-1.5: A Vision-Language-Action Model Trained using World Model- and Action-based Preference RewardsarXiv 2511.14659
Overview
- Category
- Flow·Diffusion VLA
- Type
- VLA
- Released
- 2025-11
- Organization
- Nanyang Technological University, Lambda Labs, SUTD
- Open weights
- Yes
- Parameters
- Total not disclosed; VLM 3B + action expert about 400M
Architecture
- Backbone LLM
- Qwen2.5-3B · Qwen2.5-VL-3B config
- Vision encoder
- Qwen2.5-VL ViT
- Action head
- FAST+ tokens + flow expert about 400M (hybrid)
- Chunk
- 5 steps
- Control rate
- Not disclosed
- Architecture (original summary)
- Autoregressive VLA NORA based on Qwen-2.5-VL-3B + about 400M flow matching action expert; joint optimization of FAST+ token cross-entropy and flow matching; action horizon N=5; DPO reward post-training
Training
- Pretraining accelerators
- 1 H100 node
- Pretraining duration
- 5 days
- Pretraining steps & batch
- 150k steps, global batch 512 (continued from NORA-Long 900k steps · batch 256)
- GPU/TPU-hours
- 960 (Stated · H100-hours (about 960; 1 node × 5 days))
- Fine-tuning
- Not disclosed (LIBERO·SimplerEnv·real robot·DPO)
- Data type
- Real-robot teleoperation
- Data scale
- OXE subset (same as NORA-Long, size not disclosed); 1,000 real-robot teleoperation pick-and-place episodes
- Main datasets
- Open X-Embodiment subset
- Tasks & evaluation
- LIBERO 98.0/96.0/95.4/90.5, average 95.0; SimplerEnv VM 82.8% (DPO); Galaxea A1 real-robot 9 tasks +13~46%
- Robot
- Galaxea A1 (single arm) real robot; SimplerEnv, LIBERO
Inference
- Inference device & speed
- Numbers not disclosed (only 'better inference speed')
- Inference memory
- Not disclosed
- Weights bf16 (calc.)
- Not disclosed
- Where it runs
- Not disclosed
Reinforcement learning & base
- Reinforcement learning
- Partial (Reward/preference optimization) · Uses reward-based DPO (preference optimization) in post-training. Not RL that interacts with the environment
- RL evidence
- https://arxiv.org/pdf/2511.14659
- Base control
- Fixed · No base control (Galaxea A1 single arm)
Notes
- Key features
- Attaches a flow expert to an autoregressive VLA, jointly learning FAST+ tokens and continuous actions, plus DPO reward post-training. On 9 real Galaxea A1 tasks it was 13–46% higher than baselines.
- Notes
- 960 H100-hours is the value for the 150k-step stage continued from NORA-Long. Institution corrected from the note's declare-lab per page 1 of the paper. README states discrete actions outperform continuous actions on Galaxea A1.
- Verification
- Corrections applied