Insertion Task Training Data

Claru collects insertion data using precision-instrumented assembly stations with ATI force/torque sensors recording at 1000 Hz, high-stiffness robot arms (Franka Emika, UR5e), and custom-fabricated insertion fixtures at multiple clearance levels. We record synchronized force, proprioception, and multi-view video for every insertion attempt, with full phase annotation (approach, contact, search, insert, seated) and failure mode classification (angular misalignment, lateral offset, jamming, cross-threading). Our fixture library covers common industrial connectors (USB-A/C, HDMI, Ethernet, barrel power, audio jack) and custom peg-hole geometries fabricated to client specifications. Standard delivery includes 1,000-5,000 annotated insertion attempts per geometry class, with 20-30% failure demonstrations for jam detection training, formatted for Diffusion Policy, IndustReal, TAX-Pose, or custom architectures.

Precision insertion — fitting a peg into a hole, mating a connector, starting a screw thread — is the most contact-rich manipulation task in robotics and the critical bottleneck for automated assembly. Tolerances of 0.1-0.5mm are common in electronics manufacturing (USB-C connectors have 0.3mm clearance), and success requires coordinating position control during the approach phase with force-guided compliance during the engagement phase. A robot that can reliably perform tight-tolerance insertion can automate a vast range of industrial assembly tasks worth an estimated $12 billion in annual manufacturing labor.

The core difficulty is that insertion is fundamentally a contact-dynamics problem that cannot be solved by vision alone. Once the peg contacts the hole chamfer, the robot must interpret 6-axis force/torque signals to distinguish between correct alignment (smooth insertion force), angular misalignment (moment about the contact point), lateral offset (lateral force during contact), and jamming (sudden force spike with no progress). These force signatures vary by geometry, material, friction coefficient, and clearance — making real force data from the target insertion geometries essential for policy training.

Classical insertion relies on hand-tuned spiral search strategies and compliance control parameters, which must be re-tuned for every new insertion geometry. Learned insertion policies trained on force-annotated demonstrations can generalize across geometries because they learn the underlying relationship between force patterns and corrective motions. Luo et al. (2024) demonstrated that a Diffusion Policy trained on 2,000 force-annotated insertion attempts achieves 97.3% success on peg-in-hole tasks with 0.2mm clearance, compared to 73.1% for a vision-only policy — a 24-point improvement attributable entirely to force feedback.

The data requirements for insertion are unusually demanding on temporal resolution. Force transients during initial contact last 5-20 milliseconds, and the transition from free-space motion to constrained motion occurs over 2-3 frames at standard video rates. Force/torque sensors must record at 500-1000 Hz to capture these transients, and proprioceptive data at 200+ Hz captures the compliance behavior of the robot arm during contact. Standard 30 Hz video is adequate for the approach phase but insufficient for the contact phase — high-speed cameras at 120+ Hz significantly improve insertion policy performance during the critical engagement window.

Luo et al. (2024) established the current benchmark for learned peg-in-hole insertion. Training a Diffusion Policy on 2,000 force-annotated demonstrations of peg-in-hole insertion with 0.2mm clearance, they achieved 97.3% success rate versus 73.1% for a vision-only baseline and 89.4% for a classical spiral search with optimized compliance parameters. The critical finding was that force data provides unique information about the contact state that vision cannot capture — specifically, the distinction between a peg resting on the hole chamfer (correctable misalignment) and a peg jammed against the hole wall (requires withdrawal and re-approach).

IndustReal (Tang et al., 2023) tackled the multi-geometry insertion problem. By training a single policy on 5,000 demonstrations spanning 10 different connector types (USB-A, USB-C, HDMI, Ethernet, power barrel, audio jack, and 4 industrial connectors), IndustReal achieved 91.2% average success across geometries, with the policy learning to adapt its search strategy based on the initial contact force signature. The system uses a 3-stage pipeline: vision-guided coarse alignment (to within 2mm), force-guided fine alignment (spiral search conditioned on F/T readings), and compliant insertion with progress monitoring.

TAX-Pose (Jiang et al., 2023) introduced a geometry-aware insertion approach that reasons about the relative pose between peg and hole from point cloud observations. By predicting the SE(3) transformation needed for alignment before initiating contact, TAX-Pose reduces the search phase by 60% compared to blind spiral search. Combined with a learned compliance controller for the final insertion, TAX-Pose achieved 94.7% success on novel peg-hole geometries not seen during training, demonstrating that visual pre-alignment combined with force-guided insertion is more data-efficient than either approach alone.

For deformable insertions (rubber grommets, o-rings, flexible cables into connectors), Park et al. (2024) showed that tactile sensing from GelSight sensors at the fingertips provides insertion-critical information that wrist-mounted F/T sensors miss. The contact geometry at the insertion interface — whether the deformable part is folded, compressed, or correctly aligned — is directly visible to tactile sensors but only indirectly observable through wrist force. Adding tactile data improved deformable insertion success from 71% to 93% while reducing average insertion time from 8.2 to 3.1 seconds.

Insertion data collection requires a high-stiffness robot arm (Kuka iiwa, Franka Emika, or UR5e with external F/T sensor) mounted with a calibrated 6-axis force/torque sensor at the wrist (ATI Mini45 or Nano25 for sub-Newton resolution). The F/T sensor must record at 500-1000 Hz to capture the 5-20ms contact transients that contain the alignment information. A high-resolution wrist camera provides visual pre-alignment, while overhead and side cameras capture the global approach trajectory. Proprioceptive data from joint encoders at 200+ Hz captures the arm's compliance behavior during contact.

Insertion fixtures must be precisely manufactured with known tolerances. For each insertion geometry, fabricate peg-hole pairs at 3-5 clearance levels (e.g., 0.1mm, 0.2mm, 0.5mm, 1.0mm, 2.0mm) to train policies that handle tolerance variation. Include both chamfered and unchamfered holes — chamfered holes provide a funneling effect that is common in production but absent in many research setups. Record the fixture geometry (CAD model or 3D scan), material properties (friction coefficient, stiffness), and clearance for each pair as metadata annotations.

Each insertion attempt must be annotated with: (1) insertion phase labels (free-space approach, contact detection, search/alignment, insertion, fully seated), (2) 6-axis force/torque trajectory at full sensor rate, (3) insertion depth as a function of time, (4) success/jam/failure classification with failure mode (angular misalignment, lateral offset, cross-threading for screws), and (5) the corrective motion that resolved the jam (if applicable). Failure data is as valuable as success data — policies learn to detect and recover from jams only if the training set includes the force signatures of jamming events and the corrective withdrawal-and-retry sequences.

Claru collects insertion data using precision-instrumented assembly stations with ATI force/torque sensors recording at 1000 Hz, high-stiffness robot arms, and fabricated insertion fixtures across multiple clearance levels. We record synchronized force, proprioception, and multi-view video for every attempt, with full phase annotation and failure mode classification. Standard delivery includes 1,000-5,000 annotated insertion attempts per geometry class, formatted for Diffusion Policy, IndustReal, or custom architectures.