SynthVPT

Synthetic Visual Perspective Taking: Visual Perspective Taking is the capability to infer what another agent can see and how they see it. This dataset aims at scaffolding this ability in artificial systems. Every row pairs a rendered view with a natural-language prompt and exact 6-DoF ground truth for both a human agent and everyday objects.

Synthetic rendering yields exact, noise-free pose and visibility labels at a scale and precision that annotating real images cannot reach. This frames grounded spatial reasoning as a data problem: given sufficiently precise supervision, a vision-language model can acquire the metric grounding that perspective taking requires.

v2. This release supersedes the original single-view tabletop SynthVPT. It moves from one camera and a tabletop with mugs and a humanoid to a two-agent embodied setup: a Pepper robot and a posed SMPL-X human in procedurally generated Infinigen kitchens, a diverse CanoVerse object pool, multi-tier viewpoints, and explicit perspective-taking labels.

The task

The model is a vision-language pose estimator. Given the robot's egocentric RGB and a natural-language prompt that names an entity, it regresses that entity's 6-DoF pose (translation + unit quaternion) in the robot's camera frame, for example "Assess the position and rotation of the remote control from your point of view." The entity is any object in the scene or the human, the transform column is the regression target, and the prompt selects what to localise.

Downstream, the robot composes its predicted human pose with its predicted object poses to calculate the pose of the object from the human's perspective (with metric precision). The dataset additionally carries b_sees_object (whether the human can see the object) and agt_T_obj (the object in the human's eye frame) as derived ground truth for evaluating the downstream perspective transform step.

Viewpoints

The robot's embodied camera is the deployment view; the dataset also renders feasible and free camera placements such that a model can learn pose that transfers across viewpoints rather than memorising eye height. Each row is tagged by view_role and tier:

• T0 agent_a — Pepper's embodied camera.
• T1 cam_t1_0, cam_t1_1 — feasible alternative placements a Pepper could reach.
• T2 cam_t2_0, cam_t2_1 — free extrinsics that break the eye-height-to-elevation shortcut.

Dataset structure

One row per (observer view, object query); a scene contributes several rows.

Column	Type	Meaning
image	image	the rendered RGB the model sees
prompt	string	the natural language pose request, sampled from a bank of phrasings
transform	string	4x4 object pose in the observer camera frame (the prediction target)
Tx, Ty, Tz	float	object translation in the observer frame, metres
rot_x, rot_y, rot_z, rot_w	float	object rotation in the observer frame, quaternion
semantic_class	string	object category
object_uid	string	the CanoVerse object identity
view_role, tier	string, int	which camera produced the view, and its tier
b_sees_object	bool	derived: whether the human can see the object
agt_T_obj	string	derived: 4x4 object pose in the human's eye frame
cam_T_agt	string	the human's eye pose in the observer frame
visible_px_area, edge_clipped	int, bool	object visibility in the view
symmetry_score, yaw_coherence	float	per-object rotational-symmetry and canonical-orientation quality

The prompt is drawn deterministically from a bank of verb, phrasing, and framing templates.

Generation

Scenes render in NVIDIA Omniverse Replicator. The room USD is a procedurally generated Infinigen kitchen; a Pepper robot and a posed SMPL-X human stand in the room, with objects sampled from a curated CanoVerse pool scattered on visible surfaces. Object orientation ground truth uses each object's CanoVerse canonical rotation, so poses are consistent within a category.

The kitchen is just one configurable instance. Since the environment generalises to any Infinigen scene, the same pipeline extends to any procedurally generated room or outdoor world Infinigen produces, keeping the agents, object pool, prompts, and ground truth unchanged.

Curation

This is a curated showcase subset of the full dataset; the full dataset will be released as the project matures.

Licensing

Released under a research-only license. The dataset inherits the terms of its sources:

• SMPL-X human models: non-commercial scientific research only, with registration and attribution required.
• BEDLAM / Meshcapade human skin and clothing textures: CC-BY-NC 4.0, attribution required, non-commercial.
• CanoVerse / Objaverse objects: ODC-By 1.0, with per-object attribution.
• Infinigen room geometry: the scene generator is BSD 3-Clause (Copyright (c) 2023, Princeton University), a permissive license that places no restriction on generated output; attribute the Infinigen project.

Citation

If you use SynthVPT, please cite:

@article{currie2025social,
  title={Towards Social Foundation Models: A Framework and Synthetic Dataset for Grounding Visual Perspective Taking in Robots},
  author={Currie, Joel and Piacenti, Enrico and Migno, Gioele and Gharb, Mohammad and De Tommaso, Davide and Wykowska, Agnieszka},
  year={2025}
}

@article{currie2025embodied,
  title={Towards Embodied Cognition in Robots via Spatially Grounded Synthetic Worlds},
  author={Currie, Joel and Migno, Gioele and Piacenti, Enrico and Giannaccini, Maria Elena and Bach, Patric and De Tommaso, Davide and Wykowska, Agnieszka},
  journal={arXiv preprint arXiv:2505.14366},
  year={2025}
}