Dobot
excl. VAT
Couldn't load pickup availability
Pricing Models
Compare options clearly, quickly, and without the fine print. Our recommendation is highlighted.
Flexible financing
Inclusive
Lowest price
Inclusive
Flexible terms
Inclusive
ㅤ
Switch between the tabs to view content.
Pocket Go records human-demonstrated manipulation sequences using visual and spatial motion data, thereby creating high-quality training data for learning robotic systems.
Gripping, repositioning, sorting, or tidying up can be demonstrated directly in real environments and captured as a data basis for subsequent autonomous robot tasks.
Recurring pick-and-place movements can be recorded quickly – ideal for robot learning in sorting, parts handling, and material supply.
Precise insertion, positioning, and assembly processes can be demonstrated manually and subsequently used for training robotic manipulation models.
The portable platform is suitable for research, teaching, and development projects related to embodied AI, robot learning, imitation learning, and autonomous manipulation.
From the human-operated handheld gripper to execution on a compatible DOBOT robotic arm, Pocket Go supports an end-to-end workflow for training and inference.
The Dobot Pocket Go is the ultimate mobile system for capturing high-precision, multimodal demonstration types for advanced imitation learning. Developed for AI developers, research institutions, and robotics engineers, this portable platform enables fast and flexible data generation directly in real-world application scenarios. Through the combination of high-resolution optics, visual-inertial SLAM, and an end-to-end software stack, the Pocket Go significantly accelerates the training of robust AI robot models.
Advantage for professionals: The Dobot Pocket Go solves the bottleneck in acquiring high-quality training data for AI robots. With its combination of a mobile handheld design, first-class SLAM 6D acquisition, and an end-to-end software workflow, you can drastically reduce the development effort for imitation learning and bring autonomous robot capabilities into practice faster.
Terra Robotics
https://terra-robotics.de/
mail@terra-robotics.de
Zeißstraße 10,
32584 Löhne
Germany
0.93 kg handheld kit · up to 2 kg payload capacity
The Pocket Go handheld kit weighs only 0.93 kg and enables direct manual demonstration of robotic gripping and manipulation tasks. With a gripper opening of 73.5 ± 1.5 mm, an effective gripping depth of 100 mm, and a gripping capacity of up to 2 kg, numerous everyday and industrial objects can be grasped directly in their intended operating environment.
5.3K camera · 155° field of view · integrated IMU
A GoPro Hero 12 captures the demonstration at up to 5.3K at 60 fps with a 155° field of view. Together with the integrated IMU and ORB-SLAM3, this creates a visual-inertial 6D pose tracking system that spatially records movement sequences and automatically generates usable trajectory data for robot learning.
Up to 4,000 demonstrations per day
The portable handheld concept is designed for the rapid acquisition of large demonstration volumes. Under the specified workflow conditions, up to 4,000 demonstration datasets are possible per operating day. This allows training datasets for various objects, positions, environments, and task variants to be built much more efficiently.
Data Acquisition · Training · Deployment
Pocket Go combines a handheld gripper, a robot-mountable end effector, and a complete software stack for data acquisition, calibration, trajectory processing, model training, and deployment. Trained manipulation skills can subsequently be executed on compatible DOBOT Nova and CR robot arms. A powerful mini AI computer is optionally available for local training and inference.
The DOBOT Pocket Go is a portable data collection platform for embodied AI, robot learning, and imitation learning. Human-executed grasping and manipulation tasks are recorded using camera, IMU, and 6D pose data and subsequently used as training data for robot learning models.
A user manually performs the desired task using the Pocket Go handheld gripper. During this process, the camera and IMU capture the environment and movement, while ORB-SLAM3 reconstructs the 6D spatial pose. The recorded demonstrations are subsequently processed, used to train a manipulation model, and executed on a compatible robotic arm.
Pocket Go combines high-resolution video data with inertial motion information and visual-inertial 6D pose tracking. This allows for the capture of not only camera images but also spatial motion trajectories of the demonstration, which can be used for training robotic manipulation models.
The system uses a GoPro Hero 12 with a 1/1.9-inch CMOS sensor, a 155° field of view, and HyperSmooth 6.0. Supported features include video recording at 5.3K at 60 fps, among others. The integrated IMU additionally provides motion information for visual-inertial tracking.
According to the product specification, the platform is designed for up to 4,000 demonstration data records per operating day. The actual achievable throughput depends on the task, demonstration duration, object changes, experimental setup, and organizational workflow.
The handheld kit has a standard opening of 73.5 ± 1.5 mm, an effective gripping depth of 100 mm, and is designed for a maximum gripping weight of up to 2 kg. For the robot-mounted end effector, an additional object size range of approximately 5 to 75 mm is specified.
The 0.93 kg handheld kit is guided by a human and is used to record demonstrations. The 1.58 kg end-effector kit is subsequently used on a compatible DOBOT robotic arm so that learned manipulation tasks can be executed autonomously with a comparable gripping and camera concept.
Pocket Go is designed for compatible DOBOT Nova and CR series robotic arms. The product documentation specifically lists the Nova 2, Nova 5, CR3, CR5, and CR7 for mobile setup. The required robotic arm is not part of the standard data acquisition platform and must be selected according to the project requirements.
The software workflow includes data acquisition, SLAM trajectory processing, model training, deployment, and SDK. The published system requirements specify Ubuntu 20.04 or 23.10. At least 11 GB of GPU VRAM is required for training, while at least 2 GB of GPU VRAM is provided for inference.
No. The Mini AI Computer is an optional extension for local edge training and inference. The offered configuration combines an Intel Core i9 processor with an NVIDIA RTX GPU with at least 11 GB of VRAM, enabling the training, validation, and execution of models directly on-site without a permanent cloud connection.
The DOBOT Pocket Go is a portable data acquisition platform for Embodied AI, imitation learning, and the development of autonomous manipulation skills. Instead of generating motion sequences exclusively in simulations or via complex teleoperation systems, developers can demonstrate a desired task directly with a handheld gripper. A camera, IMU, and visual-inertial tracking record the relevant movement and environmental information. These demonstrations create training data that can subsequently be used to develop robotic manipulation models.
Pocket Go is thus primarily aimed at robotics companies, AI developers, universities, and research departments that want to quickly generate high-quality data for real-world gripping, sorting, assembly, or handling tasks.
In imitation learning, a task is first demonstrated by a human. With Pocket Go, the user guides the compact handheld gripper directly through the desired motion sequence. This could be, for example, picking up an object, moving it between two containers, sorting different components, or a controlled insertion process.
The handheld kit weighs 0.93 kg and offers a gripper opening of 73.5 ± 1.5 mm as well as 100 mm of effective gripping depth. Objects up to 2 kg can be handled within the specified gripping parameters. This allows for many realistic demonstrations to be carried out without first having to program a complete robot workstation.
For companies that wish to automate such tasks later on, our page on assembly and handling offers further application examples.
A central component of the system is the GoPro Hero 12. The camera offers a field of view of 155° and recordings of up to 5.3K at 60 frames per second. HyperSmooth 6.0 supports stable video data even during dynamic motion sequences.
In addition, camera data and information from the integrated IMU are processed via ORB-SLAM3. The visual-inertial SLAM method reconstructs the spatial movement as a 6D pose with position and orientation. As a result, every demonstration produces not only video sequences but also spatially usable trajectory information for downstream training processes.
Modern Embodied AI projects often require a large number of different demonstrations. A model should not only master a task under a single ideal condition but also be able to handle changed object positions, gripping points, and environmental situations.
Pocket Go is therefore designed for high acquisition speeds. According to product specifications, the platform supports up to 4,000 demonstration datasets per working day. The actual throughput depends on the respective task and the duration of a demonstration.
Thanks to portable acquisition, variants can be recorded directly where they are relevant later on. This facilitates the creation of extensive datasets for pick-and-place, sorting, household manipulation, service tasks, or industrial assembly.
The Pocket Go concept includes a handheld gripper for human demonstration and a separate end-effector gripper kit for use on a robot arm. The robot-mounted kit weighs 1.58 kg and also has a gripper opening of 73.5 ± 1.5 mm and 100 mm of gripping depth. The specified object size range is 5 to 75 mm.
This creates a direct connection between data acquisition and subsequent robotic execution. Pocket Go is designed for compatible models of the DOBOT Nova and CR series. One possible robot system, for example, is the compact DOBOT Nova 2. An overview of other platforms is provided on our DOBOT manufacturer page.
Pocket Go does not end with recording videos. The platform concept includes a software stack for data acquisition, calibration, SLAM trajectory processing, model training, and deployment. Additionally, an SDK and a Docker-based development environment are available for individual research and integration projects.
The published system requirements mention Ubuntu 20.04 or 23.10. For model training, at least 11 GB of GPU VRAM is specified; for inference, at least 2 GB. As an optional local computing unit, a Mini AI Computer with Intel Core i9, NVIDIA RTX GPU, 64 GB DDR5, and 2 TB SSD is available.
For universities, Pocket Go opens up a practical approach to topics such as imitation learning, computer vision, SLAM, diffusion models, robot learning, and autonomous manipulation. Students can investigate the entire process from recording human demonstrations through data preparation and training to execution on a real robot.
Pocket Go thus complements classic robot programming with data-driven methods. Further robotics solutions for research and education can be found in the schools and universities section. For dual-arm experiments and more advanced robot learning scenarios, the DOBOT X-Trainer is also a related platform.
Before starting a project, the target task, objects used, desired robot arm, required amount of demonstrations, and available computing power should be defined. Especially with imitation learning, the quality and variety of the training data determine how reliably a model can handle different situations later on.
Terra Robotics assists in selecting the appropriate DOBOT platform and in setting up a suitable robotics and Embodied AI setup. For a concrete project assessment, you can submit your use case via the contact form for robotics projects.
