SSyncDev

Flora-Rewarder

SyncDev model directory · Profile updated Nov 19, 2026

Compare with another model

Flora-Rewarder is NVIDIA's 27B non-commercial reward model that scores robot episodes for success and grades each half-second of action quality.

Key specifications

Specification score
9.8
License
NVIDIA OneWay Noncommercial License

The specification score summarizes recorded features and limits. It is not a benchmark result.

Pricing conditions: Non-commercial research weights; no hosted API or price.

Inputs and outputs

Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.

Accepts

  • Video input
  • Text input
  • Image input

Produces

  • Text output

Capabilities

VisionOpen weights

What is Flora-Rewarder?

Reinforcement learning on real robots needs a reward, and success detectors written by hand break the moment the task changes. Flora-Rewarder is NVIDIA's learned alternative. It watches a robot or egocentric human video and returns either an episode-level success score or a dense, step-by-step judgement of how well the actions went.

It shipped on 15 October 2026 as part of Flora, the data engine behind GR00T-2. Its internal name is Flora-rewarder-27B.

Three kinds of reward

  1. Outcome: an integer from 1 (complete failure) to 5 (complete success), optionally after a rationale. Up to 64 frames are sampled across the episode, roughly ten per second of video.
  2. Dense: for a 16-second window, a score for every 0.5-second step, where 1 means failure, 2 suboptimal and 3 normal, plus the worst score in the window and rationales for anything non-normal.
  3. Rubric: a reasoning trace followed by one score per rubric step.

Outcome rewards are meant for monitoring deployments; dense rewards for RL post-training.

Publisher-reported results

Benchmark Metric Score
RoboRewardBench MAE on the 1–5 scale, x100 40.0
RobometerBench (OOD policy ranking) Kendall's tau-a, x100 72.0

Lower is better for the first, higher for the second.

How it was trained

Starting from Qwen3.5-27B, every parameter was fine-tuned, the vision encoder included. Each sampled time step can be a mosaic of up to four camera views, so wrist and scene cameras are judged together.

Before you wire it into a training loop

Reward models get exploited. Run a policy against it for a few thousand steps on a simple task and inspect the episodes it scores highest. If they include near-misses that look right from one camera, add views or tighten the rubric. Also measure agreement with human labels on fifty of your own episodes; the published MAE of 0.40 points on a five-point scale is a fair baseline to hold it to.

Licence: NVIDIA OneWay Noncommercial License. It is not intended to be a robot's only safety monitor, and it is not for production use.

Companions in Flora are the captioner and the temporal segmenter. For general video-language models, browse the directory.

How to evaluate Flora-Rewarder

Compare the model on the work you need to complete.

Tasks and inputs

Flora-Rewarder is listed for Video classification, Video-text-to-text. It accepts video, text, image and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.

Capacity and cost

Measure a completed task using the prompts, documents and tools your application will use. Include reasoning and retries in the result.

Choosing a deployment

The weights are downloadable under NVIDIA OneWay Noncommercial License. Check runtime support and size the complete model and working context for your expected concurrency.

Frequently asked questions

Outcome scores from 1 to 5 per episode, dense per-0.5-second action-quality scores over 16-second windows, and rubric-based scores with reasoning.

These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.

Tencent

WeMM-Video

WeMM-Video is Tencent's open Chinese-English omni-modal family (2B to 35B-A3B) that reasons over video frames and their soundtrack and answers in text.

262.14K context

NVIDIA

Flora-Captioner

Flora-Captioner is NVIDIA's 27B non-commercial video model that writes one action instruction for a chosen interval of a robot manipulation video.

— context

perceptron

Perceptron Mk1.5

Perceptron Mk1.5 is an embodied reasoning model for robots, drones and glasses: video, image and audio in; text plus boxes, points and tracks out.

32.77K context

allenai

Molmo2-4B

Molmo2-4B is Ai2's open 4B vision-language model for video, multi-image and image understanding, with pointing and object tracking built in.

— context

Ai2 (Allen Institute for AI)

allenai/MolmoPoint-Vid-4B

allenai/MolmoPoint-Vid-4B is a source-linked Hugging Face repository assigned to Video-text-to-text. Its recorded task interface accepts video, text and produces text.

— context

chenjoya

chenjoya/videollm-online-8b-v1plus

chenjoya/videollm-online-8b-v1plus is a source-linked Hugging Face repository assigned to Video-text-to-text. Its recorded task interface accepts video, text and produces text.

— context