KUKA Grasping Dataset Alternative: Beyond Bin Picking for Production Robotics

The KUKA grasping dataset was a pioneering contribution to large-scale robot learning, demonstrating that self-supervised data collection at scale could train highly effective grasping policies. As part of OXE, it contributes valuable grasping knowledge to cross-embodiment pretraining. However, production manipulation requires far more than grasping: precise placement, assembly, tool use, force-controlled insertion, and multi-step task execution are all common production requirements that the KUKA dataset does not address. Claru provides the multi-task, multi-modal demonstration data that production systems need. Our expert teleoperators collect demonstrations across your complete task repertoire on your specific robot, in your deployment environment, with the full sensor stack (RGB, depth, force/torque, proprioception, tactile) that modern manipulation architectures leverage. Teams that pretrain on OXE for grasping priors and fine-tune on Claru data for deployment-specific skills achieve the combination of broad capability and task-specific reliability that production demands.

The KUKA grasping dataset is a large-scale robotic grasping dataset collected by Google Brain, primarily associated with two landmark papers: 'Learning Hand-Eye Coordination for Robotic Grasping with Deep Learning and Large-Scale Data Collection' by Sergey Levine, Peter Pastor, Alex Krizhevsky, Julian Ibarz, and Deirdre Quillen (IJRR 2018, arXiv 1603.02199) and 'QT-Opt: Scalable Deep Reinforcement Learning for Vision-Based Robotic Manipulation' by Dmitry Kalashnikov, Alex Irpan, Peter Pastor, and colleagues (CoRL 2018, arXiv 1806.10293). The dataset was collected across two phases using fleets of 7-14 KUKA IIWA robotic arms operating continuously, accumulating over 800,000 grasp attempts in the first phase and over 900,000 in the second -- totaling approximately 580,000 curated episodes commonly referenced in the OXE collection.

The data collection approach was revolutionary for its time: robots operated autonomously around the clock, attempting grasps on bins filled with diverse small objects. An over-the-shoulder monocular RGB camera captured visual observations, while the grasp outcome (success or failure) was determined by checking whether the gripper held an object after lifting. The initial Levine et al. work trained a convolutional neural network to predict grasp success probability from camera images and proposed gripper motions, achieving closed-loop grasping that improved over time as the robot collected more data.

QT-Opt built on this foundation with a scalable deep reinforcement learning framework that used the accumulated grasping data to train a Q-function predicting long-horizon grasp success. The result was a system that achieved 96% grasp success on previously unseen objects -- a landmark result that demonstrated robots could learn general-purpose grasping through large-scale self-supervised data collection without human demonstration or hand-engineered grasp planners. QT-Opt also discovered emergent behaviors: the system learned to regrasp objects after initial failures, probe objects to find effective grasp points, and reposition items through non-prehensile pre-grasp manipulations.

The dataset is included in the Open X-Embodiment (OXE) collection under an Apache 2.0 license. It contains RGB images, gripper state (position and opening width), and binary grasp outcome labels for each attempt. The data is stored in the RLDS format on TensorFlow Datasets. Despite its pioneering role in robot learning, the dataset's single-skill focus (grasping only), limited sensor modalities (RGB and gripper state only), and specific hardware configuration (KUKA IIWA arms with parallel-jaw grippers in a structured lab setup) constrain its applicability for production systems that require diverse manipulation skills.

The most significant limitation is that the KUKA dataset covers exactly one manipulation skill: grasping. Production robots need to grasp, place, pour, screw, insert, push, pull, assemble, wipe, cut, and perform dozens of other manipulation primitives. A policy trained exclusively on KUKA grasping data can pick objects up but cannot place them precisely, cannot perform assembly operations, cannot use tools, and cannot execute the multi-step task sequences that production workflows demand. Supplementing the dataset with other OXE subsets partially addresses this, but the KUKA data itself provides no signal for non-grasping skills.

The sensor modality coverage is minimal. The dataset includes monocular RGB images from an over-the-shoulder camera and gripper state (position and opening width), but no depth data, no force/torque measurements, no full proprioceptive state (joint positions, velocities, torques), and no tactile sensing. The binary grasp outcome label (object lifted or not) is the only quality signal. Modern manipulation systems increasingly rely on multi-modal sensing -- force feedback for delicate grasps, depth for collision avoidance, tactile for slip detection -- and the KUKA dataset provides none of this information.

The data collection method (self-supervised autonomous exploration) produces a fundamentally different data distribution than expert teleoperation. The KUKA dataset contains a high proportion of failed grasps and suboptimal trajectories, because the robot was learning through trial and error. While this is appropriate for reinforcement learning, it is less useful for imitation learning approaches (behavioral cloning, diffusion policies) that expect demonstrations of successful, expert-quality task execution. Many modern robot learning architectures are designed around expert demonstrations rather than exploration data.

The environment is a highly structured lab setup: KUKA IIWA arms mounted at fixed positions, bins filled with small rigid objects placed directly in front of the gripper, consistent overhead lighting, and uniform backgrounds. Production environments are far more variable: different lighting conditions, cluttered backgrounds, objects of varying size and material, and workspaces that change configuration over time. A policy trained in the KUKA lab setup faces a substantial domain gap when deployed in a real warehouse, factory, kitchen, or retail environment.

The KUKA grasping dataset is valuable for research on grasping-specific methods: grasp prediction, grasp quality estimation, closed-loop grasping strategies, and reinforcement learning for manipulation. Its scale (580K+ attempts) makes it one of the largest single-skill datasets available, and the binary success labels enable straightforward reward shaping for RL approaches. If your research specifically focuses on improving grasping algorithms, the KUKA dataset provides a standard, large-scale training and evaluation resource.

As part of the OXE collection, the KUKA dataset contributes grasping knowledge to cross-embodiment pretraining. Policies pretrained on OXE benefit from the KUKA subset's large number of grasping attempts alongside other datasets' manipulation skills. The Apache 2.0 license makes this pretraining use commercially permissible.

Switch to commercial data when your production system needs more than just grasping. As soon as your robot must place objects precisely, assemble components, use tools, pour liquids, or execute multi-step workflows, you need demonstration data covering those skills. Claru provides multi-task demonstration data with the full spectrum of manipulation skills your deployment requires, collected on your robot platform in your environment.

The recommended pipeline for production grasping systems is: use OXE (including KUKA data) for pretraining broad manipulation priors, then fine-tune on Claru data for your specific robot, gripper, objects, and environment. For production systems requiring skills beyond grasping, Claru provides the multi-task data that the KUKA dataset fundamentally cannot.

Claru extends beyond the KUKA dataset's single-skill, single-modality paradigm to provide the full data pipeline that production manipulation systems require. Where the KUKA dataset captures 580,000 grasping attempts with RGB and binary outcomes, Claru provides expert-quality demonstrations across your complete task repertoire -- grasping, placing, assembly, tool use, pouring, insertion, and any other manipulation skills your deployment demands -- with the multi-modal sensor coverage that modern architectures leverage.

Every Claru demonstration is collected by a trained teleoperator executing the task as an expert, producing the high-quality demonstration trajectories that imitation learning architectures (behavioral cloning, diffusion policies, action chunking) are designed to learn from. Unlike the KUKA dataset's trial-and-error exploration data, Claru demonstrations represent successful, efficient task execution with smooth trajectories and appropriate force application -- the kind of data that minimizes training noise and maximizes policy performance.

Claru's sensor coverage includes everything the KUKA dataset lacks: full-resolution RGB and depth imagery from your robot's actual camera configuration, 6-axis force/torque measurements at the wrist (essential for delicate grasps, insertion tasks, and force-controlled assembly), complete proprioceptive state recording at your control frequency, and optional tactile sensor data. These modalities are precisely synchronized, providing the rich sensory context that enables policies to handle the complexity of real-world manipulation.

Data is collected on your specific robot platform -- whether that is a KUKA IIWA, Franka Emika, Universal Robots, xArm, or any other arm -- with your specific end-effector, grasping your specific objects in your actual deployment environment. This eliminates embodiment gap, gripper gap, object distribution gap, and environment gap simultaneously. All data ships with a commercial license, delivered in RLDS, HDF5, zarr, and LeRobot formats.