AgileX PiKA System: The Complete Data Collection Setup for Embodied AI (2026)

What Is AgileX PiKA?

AgileX PiKA is a family of hardware devices from AgileX Robotics (Dongguan, China) designed specifically for embodied AI data collection. The core problem PiKA addresses is that collecting high-quality manipulation demonstration data traditionally requires a physical robot arm — the operator teleoperates the arm, which constrains where you can collect data, how quickly you can set up new environments, and how many demonstrations you can gather per unit time.

PiKA takes a different approach: the human operator becomes the data source. A hand-held device (PiKA Sense-H) captures the operator's hand trajectory, camera observations, and spatial position in real time. Positioning beacons (PiKA Station-H) anchor the coordinate system so demonstrations from different sessions and locations can be registered to a common spatial frame. The resulting dataset captures human manipulation behavior directly, without requiring a robot to be physically present during collection.

This approach has precedent in academic research — the UMI (Universal Manipulation Interface) paper from Chi et al. (2024) demonstrated that hand-held gripper devices can produce data that transfers to robot arms. AgileX's contribution with PiKA is to build a production-ready, commercially supported version of this concept with dedicated positioning infrastructure, an optional compute backpack for on-site processing, and factory integration with their robot arm lineup.

PiKA hardware is available at SVRC and ships from Dongguan with a 3-5 week lead time. See our PiKA Sense-H store listing and the broader AgileX hardware collection at SVRC.

Full System Breakdown

The PiKA system consists of five distinct hardware components. Each is sold separately, which allows teams to buy only what they need for their specific use case. Here is a complete breakdown of each device.

PiKA Sense-H: The Handheld Data Collection Device

PiKA Sense-H is the core collection device — the item the operator holds in their hand during demonstrations. It is a handheld unit that integrates multiple sensors into a single ergonomic form factor.

Sense-H captures the operator's end-effector trajectory (the path of the human hand through space) along with synchronized camera observations. The 120Hz positioning rate provides enough temporal resolution to capture fast manipulation movements without motion blur in the trajectory data. The depth camera enables downstream 3D reconstruction of the manipulation workspace, which is useful for sim-to-real transfer and dataset visualization.

PiKA Gripper-H: Arm-Mounted Data Collection Gripper

PiKA Gripper-H is a 2-finger gripper with integrated cameras that mounts on a robot arm. Rather than the operator holding a handheld device, Gripper-H is attached to the robot arm's wrist and provides a close-up visual observation stream from the gripper's perspective, synchronized with the arm's joint state data.

PiKA Station-H: Spatial Positioning Beacon

PiKA Station-H is the UWB (Ultra-Wideband) positioning beacon that anchors the spatial coordinate system for PiKA Sense-H demonstrations. Station-H units are placed in known positions in the demonstration environment; they transmit UWB pulses that Sense-H uses to compute its 3D position at 120Hz. A minimum of 2 Station-H units are required to establish a baseline coordinate frame; 3 or more are recommended for more robust 3D positioning coverage.

The 30-hour battery life allows Station-H units to run through a full day of data collection and overnight without recharging. This is relevant for teams running continuous multi-shift collection campaigns. For indoor environments with good UWB signal propagation, 2 Station-H units covering a ~3m x 3m workspace are sufficient for the majority of tabletop and countertop manipulation scenarios.

PiKA Ego-H: First-Person Camera

PiKA Ego-H is a first-person camera worn or mounted on the operator during data collection. It captures the operator's visual perspective — what the human sees while performing the manipulation task. This ego-centric view is a key observation modality for training embodied AI policies that need to generalize to a robot's own perspective, particularly for tasks where the robot will eventually operate in environments similar to where the human demonstrated.

PiKA OS-H: The Compute Backpack

PiKA OS-H is a wearable compute unit — essentially a laptop-class computer in a backpack form factor — that runs all PiKA data processing on the operator's body during collection. Rather than streaming raw sensor data to an external workstation over Wi-Fi, OS-H processes and records everything locally. This eliminates latency issues and network bandwidth requirements that would otherwise limit collection throughput in multi-person campaigns or off-site environments.

The 3-hour runtime matches a standard data collection shift. For longer campaigns, operators can swap battery packs or plug into wall power between recording sessions. The 2TB local storage holds roughly 20-40 hours of synchronized multi-camera + sensor data at typical collection quality settings before requiring offload.

How Data Collection Works

A complete PiKA data collection session using the full system (Sense-H + Station-H x2 + Ego-H + OS-H) follows this workflow:

1. Environment setup: Place 2-3 Station-H beacons in known positions around the demonstration workspace. The beacons self-calibrate their relative positions automatically once powered on.
2. Operator equipment: Put on the OS-H backpack, mount Ego-H at head or chest level, pick up Sense-H in the dominant hand.
3. System initialization: OS-H software connects to all devices, verifies sensor streams, and establishes the spatial coordinate frame from the Station-H network. Takes approximately 2-3 minutes.
4. Recording: Press the record button on Sense-H. All sensors begin synchronized recording: Sense-H trajectory at 120Hz, depth and fisheye video, Ego-H video, and environmental metadata.
5. Demonstration: Operator performs the manipulation task naturally. Sense-H captures the hand trajectory; cameras capture visual context from multiple perspectives.
6. Episode segmentation: A button press or voice command marks episode boundaries. The OS-H software writes a complete HDF5 episode file per demonstration.
7. Data transfer: At end of session, episode files are synced from OS-H to a workstation or cloud storage for annotation, review, and training pipeline ingestion.

The dataset produced by PiKA is a set of synchronized episode files, each containing the operator's 6-DOF end-effector trajectory (in the Station-H coordinate frame), camera observations from multiple viewpoints, timestamps, and gripper state. This format is compatible with LeRobot training pipelines and can be converted to the observation/action format required by ACT or Diffusion Policy training with the provided conversion scripts.

The central value proposition of PiKA over fixed-arm teleoperation is the ability to collect data anywhere. Because Station-H beacons set up in minutes and Sense-H requires no arm installation, a PiKA operator can collect demonstrations in a restaurant kitchen, a warehouse, a hospital room, or any other environment where building a robot arm station would be impractical. See our ALOHA vs UMI data collection guide for more context on in-the-wild vs lab-based data collection tradeoffs.

Pricing Table

All PiKA hardware ships from Dongguan, China with a 3-5 week lead time from order placement. SVRC handles import, QA inspection, and US-based support. Contact us for volume pricing if you are deploying multiple collection kits simultaneously. See the PiKA Sense-H store listing and PiKA Ego-H listing for current availability.

PiKA vs UMI vs DexCap

Three hardware approaches dominate the "in-the-wild" manipulation data collection space: AgileX PiKA, UMI (Universal Manipulation Interface), and DexCap. Here is a direct comparison:

The fundamental tradeoff: PiKA is dramatically more expensive than DIY UMI or DexCap builds, but provides production-grade reliability, dedicated spatial positioning infrastructure, on-device compute, and commercial support. For a university lab or individual researcher, a DIY UMI build is often the right starting point. For a company planning to run a professional data collection operation with multiple operators over extended campaigns, PiKA's infrastructure investment pays off quickly in reduced downtime and higher data quality consistency.

DexCap targets a different use case entirely — full dexterous hand motion capture rather than gripper-style demonstrations. If your target robot has a dexterous hand rather than a parallel gripper, DexCap-style capture is more appropriate than either PiKA or UMI. See our dexterous hand guide for context on hand hardware options.

Where to Buy

All AgileX PiKA components are available at SVRC. We are an authorized AgileX distributor in North America, handling import, QA, and local support for all AgileX hardware. PiKA hardware ships from Dongguan with a 3-5 week lead time; in-stock units ship within 1-2 business days.

Frequently Asked Questions

What is AgileX PiKA?

AgileX PiKA is a modular hardware system from AgileX Robotics for embodied AI data collection. The system includes five components — PiKA Sense-H (handheld sensor), Gripper-H (arm-mounted gripper), Station-H (UWB positioning beacon), Ego-H (first-person camera), and OS-H (compute backpack) — that together enable human operators to collect manipulation demonstrations without requiring a fixed robot arm to be present during collection.

What is the price of AgileX PiKA?

Component prices at SVRC: PiKA Sense-H $9,998; PiKA Gripper-H $10,998; PiKA Station-H $6,998 each (minimum 2 required); PiKA Ego-H $5,998; PiKA OS-H $59,998. A minimal mobile collection kit (Sense-H + 2x Station-H) starts at approximately $23,994. A full system with all five components and 2 Station-H units is approximately $100,988. Ships from Dongguan, 3-5 week lead time.

Is AgileX PiKA compatible with ROS2 and LeRobot?

Yes. PiKA hardware is ROS2 compatible with Ubuntu 22.04 and ROS2 Humble pre-installed on the OS-H backpack. Data recorded by PiKA can be exported in LeRobot dataset format for training with Hugging Face's framework, including ACT, Diffusion Policy, and other VLA architectures.

How does AgileX PiKA compare to UMI and DexCap?

PiKA provides production-grade reliability with dedicated UWB positioning (Station-H) for accurate spatial anchoring, an optional compute backpack (OS-H) for on-device processing, and commercial support from AgileX and SVRC. UMI is a DIY-friendly academic design using a GoPro-mounted gripper with visual-inertial odometry — lower cost (~$500-2,000 DIY) but requires significant self-setup. DexCap captures full dexterous hand motion for finger-level manipulation, a different use case than PiKA's gripper-focused approach. Choose PiKA for professional data operations; consider UMI for budget-constrained research starting points.