leopoldmaillard/sceneteract-grpo

SceneTeract GRPO Training Set

Action-level feasibility samples for post-training a VLM against a geometric verifier. Each row is one atomic interaction — an image, a prompt, and a label that was measured rather than annotated — ready to drop into TRL's GRPOTrainer.

8,073 samples over 1,132 3D-FRONT living rooms and dining rooms and three agent profiles.

from datasets import load_dataset

ds = load_dataset("leopoldmaillard/sceneteract-grpo")
ds["train"]   # 6,473 samples /  905 scenes
ds["test"]    # 1,600 samples /  227 scenes

Columns

column	type
image	Image	top-down view of the room with the target object in a red bounding box
messages	list⟨struct⟩	system + user turns, with an image placeholder, in TRL's chat format
ground_truth	string	"True" / "False" — a string, not a bool, as TRL reward functions compare strings
failure_property	string	the check that failed, or the string "None" when the action is feasible
action	string	the atomic action, e.g. open, sit_on, take_out_of
agent_id	string	adult, child, adult_wheelchair
scene_id	string	3D-FRONT scene, as preprocessed by PhyScene

Both ground_truth and failure_property are strings including their "empty" cases — "False" and "None" are values, not nulls. Compare them as strings.

Splits

Scene-disjoint, 80/20. The same partition ships in sceneteract-traces, so a model trained here can be evaluated on the traces' test split without leakage.

Training

ground_truth is the reward signal: a rollout is correct when its predicted feasibility matches the verifier's measurement. The reward functions used — format, correctness, and spatial shaping — ship with the code:

pip install "sceneteract[grpo]"
sceneteract-grpo train --dataset <path> --model Qwen/Qwen3-VL-4B-Instruct

See docs/usecase_c_grpo.md for reward shaping and hyperparameters.

Licence and attribution

Released under *CC BY-NC 4.0.

If you use this dataset, please cite those works alongside SceneTeract:

• 3D-FRONT — Fu et al., 3D-FRONT: 3D Furnished Rooms with layOuts and semaNTics, ICCV 2021
• 3D-FUTURE — Fu et al., 3D-FUTURE: 3D Furniture Shape with TextURE, IJCV 2021
• PhyScene — Yang et al., PhyScene: Physically Interactable 3D Scene Synthesis for Embodied AI, CVPR 2024

Citation

@misc{maillard2026sceneteract,
      title={{SceneTeract: Probing and Improving Agent-Aware Activity Reasoning in 3D Indoor Scenes}}, 
      author={Léopold Maillard and Francis Engelmann and Tom Durand and Boxiao Pan and Yang You and Leonidas Guibas and Maks Ovsjanikov},
      year={2026},
      eprint={2603.29798},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2603.29798}, 
}