Door & Cabinet Manipulation Training Data

Claru collects articulated object manipulation data across real homes, offices, and commercial buildings through our distributed collection network, providing the hardware diversity no single lab can achieve. Each site is surveyed and cataloged for handle types, hinge mechanisms, and opening configurations. Our teleoperation rigs capture full interaction episodes with 200 Hz force/torque sensing, multi-view RGB-D, and detailed hardware metadata. Annotations include handle localization, articulation axis estimation, force profile phase segmentation, and opening angle trajectories. We deliver datasets covering 50+ hardware configurations in RLDS, HDF5, or custom formats with full sensor calibration and hardware-stratified train/val/test splits. Typical campaigns across 10+ collection sites produce 5,000-10,000 annotated episodes within 4-6 weeks.

Door and cabinet manipulation is the task of operating articulated objects — doors with hinges, drawers on rails, cabinets with various handle types, refrigerators with seals, and oven doors with counterbalance springs. Every home, office, and commercial building contains dozens of these objects, making articulated manipulation a prerequisite for any robot operating in human environments. The fundamental challenge is hardware diversity: a lever handle requires a downward press, a knob requires a grasp and twist, a pull handle requires a firm grip, and a push plate requires whole-hand contact — each with different force profiles and approach trajectories.

The manipulation challenge compounds because articulated objects have constrained motion. A cabinet door swings on a hinge with a fixed axis of rotation, but the robot does not know that axis a priori. The robot must discover the articulation model (revolute joint for doors, prismatic joint for drawers) from initial contact and then adapt its trajectory to follow the constraint. Where2Act (Mo et al., 2021) demonstrated that learning where to interact on an articulated object is as important as learning how to interact — predicting actionable contact points from visual observation requires training data spanning the enormous variety of handle designs and mounting positions found in the real world.

Commercial interest is driven by the home robot market (every service robot must open doors and cabinets), retail automation (shelf restocking behind cabinet doors), and healthcare (medication carts, supply closets). The data bottleneck is particularly acute because each hardware configuration (handle type, hinge type, mounting height, opening direction) requires dedicated demonstrations. A kitchen alone may contain 15-20 distinct cabinet configurations, and a production-grade policy must handle all of them.

The problem compounds when you consider that even the same handle type behaves differently depending on mounting context. A lever handle at waist height requires a different approach trajectory than the same lever handle overhead. A drawer pull at floor level requires the robot to crouch, changing the entire reachability and force application geometry. Height-indexed data — demonstrations stratified by the vertical position of the handle relative to the robot's base — is a critical and often overlooked dimension. Without it, policies trained on countertop-height cabinets fail systematically on overhead cabinets and under-counter drawers, despite the underlying handle mechanism being identical.

The field has evolved from model-based approaches (estimating articulation parameters from point clouds, then planning compliant trajectories) to data-driven end-to-end policies. Where2Act (Mo et al., 2021) trains a point cloud-based model to predict per-point action scores — indicating where on an object to push, pull, or grasp — and achieves 70% success on unseen articulated objects in simulation. FlowBot3D (Eisner et al., 2022) improves on this by predicting 3D flow fields that indicate the desired motion direction at each point, allowing the robot to discover the articulation axis without explicit model estimation.

Real-world results remain challenging. RT-2 (Brohan et al., 2023) can open drawers and cabinets from language instructions but requires the handle to be clearly visible and accessible. Diffusion Policy (Chi et al., 2023) achieves 85-90% success on opening specific drawer types from 200 demonstrations, but success drops to 60-70% on unseen drawer hardware. The gap is primarily in handle detection and initial grasp pose — once the robot has a firm grasp and the articulation axis is discovered, compliant trajectory following is reliable.

The emerging approach uses articulation-aware foundation models. Rather than training separate policies per hardware type, these models learn a shared representation of articulated objects from large-scale data and predict articulation parameters (joint type, axis, range) alongside manipulation actions. This requires diverse training data spanning dozens of hardware types — exactly the kind of data that single-lab collection struggles to produce but distributed collection networks can achieve.

Mobile manipulation adds another layer: the robot must coordinate base positioning with arm motion during articulation. Opening a full-size door requires stepping backward while pulling, and accessing a low cabinet requires the base to position the arm within reach of a handle near the floor. Mobile ALOHA (Fu et al., 2024) demonstrated that mobile door opening from 50 demonstrations is feasible, but the base coordination patterns are highly sensitive to the door's swing direction and resistance. Data for mobile articulated manipulation must capture the coupled base-arm coordination, including recovery behaviors when the base position drifts during a pull or push.

Articulated object data collection requires physical diversity that cannot be replicated with a single test fixture. Claru collects across real kitchens, offices, bathrooms, and commercial spaces where the natural variation in hardware provides the distribution our policies need. Each collection site is surveyed to catalog the articulated objects: handle type (lever, knob, pull, push plate, recessed), hinge type (butt, European cup, piano), opening direction (left, right, pull-out), and any resistance mechanisms (magnetic catches, soft-close dampers, spring returns).

The demonstration protocol captures the full interaction: approach (moving to the handle), grasp (making contact and securing grip), actuation (opening/closing along the articulated path), and release (disengaging from the handle). Force/torque data at 200 Hz is critical because it captures the initial pull force to overcome catches, the steady-state force during articulation, and the changing force profile as soft-close mechanisms engage. Annotations include handle bounding box, articulation axis estimation (from the trajectory), opening angle achieved, and force profile segmentation into phases (approach, break-free, sweep, decelerate).

Our collection protocol ensures a minimum of 10 distinct hardware configurations per collection site, with 50-100 demonstrations per configuration across 3+ operators. This produces 500-1,000 episodes per site. We target 10+ sites per campaign to achieve the 5,000-10,000 episode datasets needed for cross-hardware generalization. Each episode is tagged with detailed hardware metadata (handle manufacturer and model when identifiable) to enable hardware-conditioned policy training.

For mobile manipulation scenarios, demonstrations capture the full approach-grasp-actuate-release sequence including base repositioning. The robot's base trajectory is logged alongside arm actions, enabling training of coordinated base-arm policies. We annotate the required clear space for door swing, the optimal base standoff distance per door type, and any coordination events where the base must move during articulation. This mobile coordination data is essential for home and commercial robot deployments where the robot must navigate to, operate, and navigate away from articulated objects as part of longer task sequences.