Crobotia Robotics

Crobotia is a robot foundation-model team building verified, multimodal data for real-world manipulation. We capture demonstrations in everyday environments, refine them into training-ready batches, and connect each delivery to a training recipe, checkpoint, and real-robot evaluation.

Our focus: robot learning · embodied AI · multimodal demonstrations · neural visual-inertial odometry · task and affordance annotation

Real environments. Instrumented demonstrations.

Robots need to learn from the homes and objects they will encounter. Our collection pipeline captures everyday manipulation with the lighting, clutter, and variation of real rooms.

Actual Crobotia capture: garment held by a gripper in a bedroom

Actual Crobotia capture: folding clothes on a patterned bedspread

Actual Crobotia capture: a bedroom with a window and desk fan

_{Real capture frames from crobotia.com. The point-cloud banner above is an AI-generated brand illustration.}

A closed loop from capture to evaluation

Capture in real environments → refine synchronized and labeled batches → train and evaluate on real robots → use failures to guide the next collection.

Evaluation informs the next collection round. We use policy behavior on real robots to identify what to capture next, keeping collection, training, and validation connected.

What comes with a delivered batch

Refined demonstrations: idle and drifting segments removed, with aligned streams and per-session metadata.
Calibration and geometry: per-device intrinsics and extrinsics, metric trajectories, and 3D scene context.
Reviewed annotations: tasks, ordered action steps, task progress, and grounded affordances, proposed by a vision-language model and checked by a person.
Training and evaluation artifacts: the training recipe, checkpoint trained on the batch, and its real-robot evaluation report.

One time base. Multiple modalities.

Modality	Capture specification
Vision	Dual 200° fisheye views · 4K per lens · 50 fps
Motion & geometry	Neural visual-inertial odometry · metric camera trajectories · dense colored point clouds
Actions	Gripper position and orientation at 50 Hz, with jaw angle
Inertial sensing	Calibrated accelerometer and gyroscope streams at 200 Hz
Audio	Stereo audio at 48 kHz on the shared time base
Annotations & metadata	Human-reviewed task and affordance labels, device calibration, and session provenance

Every session is checked for synchronization. See the time-base design and VIO accuracy report for technical details.

From capture hardware to training data

Instrumented capture rigs	An annotated ego view

_{Product and capture images from Crobotia. Explore the rig · Inspect a delivered data unit}

Explore & collaborate

Resource	What you can explore
Dataset preview →	Ego views, approved instructions, action steps, and grounded moments
VIO accuracy report →	Trajectory evaluation and per-recording errors
Collection & evaluation loop →	How model feedback guides data collection

We work with teams on datasets, capture rigs, and pipeline deployment within their own infrastructure. Tell us what you are training and which tasks your policy needs to improve.

Request access → · hello@crobotia.com · @crobotia_ai