SynthVPT

As robots enter the physical spaces in which people live and work, an essential step towards shared understanding with their human interaction partners is Visual Perspective Taking (VPT): the capacity to represent what another agent can see, and how a shared environment appears from their distinct frame of reference. SynthVPT is a procedurally generated synthetic dataset providing exact metric poses for human agents (eye position and gaze) and objects across 140 categories (in their canonical frames of reference). Each rendered view is paired with a natural-language request naming one entity and with that entity's exact annotated pose in the reference frame of the camera that produced the render. Each frame is rendered from a humanoid robot's perspective, from auxiliary viewpoints, and from the human's own eyes. The full dataset is yet to be released publicly.

The SynthVPT generation pipeline. The scene's subcomponents are first converted to USD, then composed into a single scene graph, which Replicator re-scatters, resamples and rerenders on every frame. Six cameras render each frame; the two shown are Pepper's own camera, and the camera rendering the human's viewpoint. One view paired with one entity, its prompt and that entity's annotated pose forms one sample of the dataset. Annotated reference frames are drawn with z up, the green ray on the human marking the gaze direction −y.

Perspective taking

Empirical research in humans has long decomposed VPT into two distinct capabilities, each representing a level of perspective taking that can support progressively more sophisticated human-robot collaboration:

• VPT-1: Estimate the perceptual access from another viewpoint: what can they see?
• VPT-2: Estimate the perceptual content from another viewpoint: how does it appear to them?

Both capabilities can be emulated using the same underlying representation of the other agent (here, the human) and the referent object (the object a prompt names): the position and orientation of each relative to one's own frame of reference; in short, they require egocentric spatial estimation. Both VPT-1 and VPT-2 can therefore be derived by geometrically transforming an entity's pose into the reference frame of another agent.

Design decisions

Each scene places an observer (the Pepper robot) and a target agent (a human partner) in a procedurally generated indoor environment, with commonplace objects scattered across the available surfaces.

• Synthetic people Our pipeline integrates SMPL-X, which provides a parametric body model, in which the agent's morphology can be specified at render time by a set of shape coefficients, deforming a SMPL-X's body template such that body proportions vary continuously. The agent's body posture is sampled from VPoser, a learned pose prior trained on motion capture data and distributed with SMPL-X. The agent's appearance is composited at render time from 102 skin textures by Meshcapade (100 via BEDLAM, two sample textures) and 1,737 clothing textures from the BEDLAM texture pack, a synthetic human dataset with demonstrated success for sim-to-real transfer. In sum, the agent's body shape, posture and visual appearance are resampled for every frame, ensuring each sample of the dataset effectively presents a different person.
• Synthetic robots We chose Pepper for its humanoid form and head-mounted camera, which provide a human-like, egocentric perspective. As a mobile agent with a constantly shifting frame of reference, Pepper must localise entities from its current field of view. This is the estimation problem SynthVPT aims to advance, i.e. deriving the camera-relative poses of agents and objects from a single monocular image. Pepper also remains a commonly used human-robot interaction platform, with a public kinematic description that can be utilised in simulation.
• The human's pose The reference frame of the target agent is anchored at the agent's eyes, instead of the agent as a whole, its forward axis the direction of the agent's gaze.
• Object orientation An object's front is a tacit convention among people, rather than a property of its geometry, as such we use the object orientation annotations from CanoVerse, a released library of Objaverse 3D assets annotated with per-category canonical reference frames (as an example a camera faces the direction its lens points, and a boot's forward is dictated by how you would wear it), in which a human annotator chooses which of five candidate poses matches the category.
• Assets Of the original 31,980 asset subset of CanoVerse, 8,201 assets across 593 categories were retained after quality and scale filtering.
• Scenes The base scene is generated with Infinigen Indoors, which produces indoor 3D environments in which room geometry, materials, furniture and appliances are derived algorithmically instead of using a fixed library of models. For this dataset, we held the environment fixed as a single kitchen acting as a proof of concept, with the pipeline designed to scale to infinite room variations across six room archetypes. The illumination configuration is driven at render time: rendering variations of the sky, the sunlight through the kitchen's windows, and the room's own lamps.
• Prompts Individual prompts are generated from a bank of 500 templates, to reduce the likelihood of a model overfitting to a limited number of prompts. The prompt bank crosses fourteen imperative verbs (e.g. Find, Estimate, Report, …) with seven ways of naming the pose (e.g. position and orientation, 6-DoF pose, spatial pose, …), and five sentence frames (e.g. Find the pose of the mug, Find the mug's pose, Find the pose of the mug relative to you, …). In addition ten interrogative forms are included (Where is the mug and which way is it facing?, …). Three of the five sentence frames ask for the pose using embodied language, from the observer's own frame of reference (e.g. relative to you, from your point of view, or from where you are).
• Viewpoints Renders of the scene are captured from several distinct viewpoints, in three tiers ordered by how closely each corresponds to a view the deployed robot could take, plus the human's own view:
  • embodied: the forehead camera of the 3D robot model standing in the scene, reproducing the physical robot's focal length (57.2° field of view), the camera direction driven by the robot's gaze behaviour.
  • feasible: viewpoints the robot could physically take: at its eye level, on floorspace that would accommodate its base, with the head direction kept within the robot's joint limits.
  • free: positioned without the embodied constraints, sampling the camera height uniformly between 0.4 m and 2.2 m.
  • human: the synthetic human's viewpoint, with a field of view of 120°, approximating the span of human binocular vision.

This release

SynthVPT contains 240 scenes, each of 250 frames, with each frame rendered from six distinct views. We split the dataset by scene, meaning that every frame of a scene is allocated to the same dataset split (training, validation, test). The full dataset holds 702,190 samples across 140 object categories, counting every view except the human's.

This repository holds a curated subset of 1,267 samples from 61 frames across 23 scenes, covering 102 object categories: 444 from Pepper's camera (embodied), 327 feasible, 364 free, and 132 from the human's eyes (human). Rows are grouped by frame, so each moment can be read from every view in turn.

Column	Meaning
tier	the view: embodied, feasible, free or human
transform	the named entity's pose in the camera's frame (4 × 4, metres; x right, y down, z forward), also given as Tx, Ty, Tz and the quaternion rot_x, rot_y, rot_z, rot_w
cam_T_agt	the human's eye pose in the camera's frame
agt_T_obj	on object rows, the object's pose in the human's eye frame: cam_T_agt⁻¹ · transform
b_sees_object	on object rows, whether the human can see the object, read from the human's own render, so an object hidden behind furniture counts as unseen
b_in_fov	on object rows, whether the object's centre falls within the human's 120° × 120° field of view

The remaining columns are visible in the dataset viewer above.

A paper describing SynthVPT is in preparation.

Sim-to-real transfer

To illustrate what VPT judgements look like in a deployment setting, we gave a vision-language model, fine-tuned only on SynthVPT, two photographs from our lab.

Demonstration of the model's VPT-1 capability with real photographs. We provided the model (only fine-tuned in simulation) with two photos of our lab: (a) a man (@jwgcurrie) looks downwards, towards a toolbox; in (b) he directs his gaze to the robot (iCub) across the table. Shown are the photographs with each predicted position annotated, a bird's-eye view generated using the model's estimates, and an orthographic 3D view. The four text prompts below are used identically for both photographs. The man is the only target agent in the scene; the robot, screwdriver and box are referent objects. The red shaded region represents the man's predicted field of view, drawn in plan with its extent following the full 3-DoF gaze estimate, meaning that looking downwards shortens it.

Licensing

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

• SMPL-X bodies: SMPL-X Body license (CC-BY 4.0).
• BEDLAM body and clothing textures: BEDLAM license.
• Meshcapade sample skin textures: CC-BY 4.0.
• CanoVerse / Objaverse objects: ODC-By 1.0, with each object's licence listed under its object_uid in Objaverse's metadata.
• Infinigen room geometry: BSD 3-Clause.
• Pepper robot meshes, from qiBullet: CC BY-NC-ND 4.0.