leopoldmaillard/sceneteract-traces

SceneTeract Verification Traces

Geometry-grounded feasibility labels for everyday activities in 3D indoor scenes, conditioned on who is performing them.

Each row is one verification trace: a (scene, agent, activity) triplet in which a VLM decomposed the activity into atomic (action, object) steps, and a geometric engine measured whether each step is physically executable — returning a per-check verdict and the reason it failed.

3,396 traces over 1,132 3D-FRONT living rooms, dining rooms and combined living-dining rooms, three agent profiles (adult, child, adult_wheelchair), 583 distinct activities, 15,498 verified steps and 20,344 individual geometric checks.

from datasets import load_dataset

ds = load_dataset("leopoldmaillard/sceneteract-traces")
ds["test"]   #   681 traces /  227 scenes — held-out evaluation subset
ds["train"]  # 2,715 traces /  905 scenes — scenes used for GRPO post-training

What a row looks like

A row mirrors the result.json that the verifier writes:

column	type
run_id	string	{scene_id}_{agent_id}
scene_id	string	3D-FRONT scene, as preprocessed by PhyScene
room_type	string	LivingRoom, DiningRoom or LivingDiningRoom
agent_id	string	adult, child, adult_wheelchair
task	string	the activity, in natural language
planner_model	string	the VLM that produced the plan
success	bool	the label — is the whole activity executable
failure_reason	string	the first check that failed, null if feasible
failed_step_index	int32	null if feasible
total_steps	int32	steps in the plan
plan	list⟨struct⟩	object_id, semantic_action, object_category, properties
step_reports	list⟨struct⟩	per step: action, object_id, object_category, is_success, property_reports
insight	string	model-written explanation of the failure

step_reports[].property_reports[] is where the diagnostic lives: each entry has a property_name, an is_success, and a message stating the measurement.

is_Reachable    True    "Object is reachable. Required distance: 0.19m, Agent's reach: 0.70m."
has_Clearance   False   "All potential interaction zones are blocked by obstacles."

The five checks

property	what it measures
is_Navigable_To	a collision-free path exists from the agent to an interaction zone of the target
is_Reachable	the target is within the agent's reach from a navigable position
is_Interactable	the object's functional part (a handle, a knob) is reachable, not just the object
has_Clearance	enough free volume to actually perform the action — open a cabinet
is_Visible	the target is unoccluded from the agent's eye sight

Which checks an action requires is fixed by the action taxonomy: sit_on needs navigation and clearance, open additionally needs the handle to be interactable.

Splits

test is the 227 scenes held out from GRPO post-training and is the intended evaluation subset; train is the remaining 905 scenes, used for post-training. The partition is scene-disjoint, so no scene appears in both and a model trained on train can be evaluated on test without leakage.

For scene-audit statistics over the full corpus, concatenate them:

from datasets import concatenate_datasets
full = concatenate_datasets([ds["train"], ds["test"]])   # 3,396 traces

Flattening to action or check level

The nesting holds the action and property granularities; one comprehension gets you either as a flat table.

import pandas as pd

checks = pd.DataFrame([
    {"run_id": t["run_id"], "scene_id": t["scene_id"], "agent_id": t["agent_id"],
     "step": i, "action": s["action"], "object_category": s["object_category"],
     "property": c["property_name"], "passed": c["is_success"], "message": c["message"]}
    for t in full
    for i, s in enumerate(t["step_reports"])
    for c in s["property_reports"]
])                                        # 20,344 checks over 15,498 steps

# which constraint blocks wheelchair users most often?
fails = checks[(checks.agent_id == "adult_wheelchair") & ~checks.passed]
fails["property"].value_counts()

How the labels were produced

Plans were generated by a single VLM planner (google/gemini-3-flash-preview) from a rendered view of the scene; the same model resolves which interaction zone of an object an action targets. Every feasibility verdict is then measured geometrically.

insight is model-generated prose. It summarises the failure in natural language and is useful for reading traces.

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}, 
}