Pollen Robotics open-sources Grabette, a handheld rig for robot data
Hugging Face
- Robotics
- Open source
- Data
- #grabette
- #lerobot
- #manipulation
Pollen Robotics published Grabette, an open handheld gripper that records manipulation demonstrations without a robot and converts them into LeRobot datasets, plus Gripette, a motorized gripper twin for robot arms. The post attributes its design to Stanford's UMI and demonstrates the pipeline with a Diffusion Policy example; the full hardware and training details were not independently reviewed.
What Pollen released and how it works
Pollen Robotics published Grabette, described as an open, low-cost handheld system for recording robot-manipulation demonstrations without a robot. The device carries a wide fisheye camera that gives the policy a context-rich, wrist-camera-style view, and an RGBD camera plus IMU that the post says handles 6-DoF tracking, along with gripper encoders. Pressing a button records those streams on a Raspberry Pi on a shared clock; a browser dashboard then runs SLAM with RTAB-MAP, checks for jumps or tracking loss, converts episodes to LeRobot format and uploads a dataset to Hugging Face. A motorized counterpart, Gripette, is the end-effector for a real or simulated arm. The post credits Stanford's UMI as the direct inspiration and states a ~490€ build-of-materials cost for Grabette and ~120€ for Gripette. It also outlines an example that trains a Diffusion Policy on 200 demonstrations and evaluates it on a 7-DoF OpenArm, while stressing that this illustrates the pipeline rather than being the product.
Where a lab might try it
A robotics research engineer whose arm has limited teleoperation hardware could use Grabette to collect demonstrations by hand in workspaces a robot cannot easily reach. Because episodes are stored as camera-local 6-DoF cartesian pose plus gripper state and exported as LeRobot datasets, a team whose training stack already reads LeRobot could try recording a small task battery without writing a bespoke capture tool. The matching Gripette gripper on the arm would be a prerequisite for closing the loop on hardware rather than simulation.
A small, measurable trial
A team could pick one short manipulation task, record a modest number of episodes with Grabette, and compare a policy trained on them against a baseline trained on whatever data it already uses, holding the training stack fixed. Before training, check the dashboard's SLAM validation for tracking loss and jumps and discard flagged episodes rather than training through them; a human should review episode quality. Success could be judged on task success rate and failure modes during a fixed evaluation protocol on the team's own arm, not just on training loss. A reasonable failure check is whether demonstrations transfer to the target arm at all, since the post requires the matching Gripette gripper.
Trade-offs and unknowns
Handheld capture removes teleoperation logistics but introduces its own risks: tracking quality depends on the RGBD camera and SLAM step, and the post does not quantify how often episodes fail validation. Costs cited are publisher-stated bill-of-materials figures, not quotes for a working unit. The example was run with a Diffusion Policy on the OpenArm; whether the same data format and pipeline yield useful policies on other arms or with other methods is not evidenced in the excerpt. This may not suit teams needing force or tactile signals, or those whose tasks depend on contacts a handheld rig cannot reproduce. Read the full post and hardware repositories before committing budget.
Context and limitations
- Based on a short excerpt of the post; the full article, hardware repos and code were not reviewed.
- Pollen's reported build costs and example results are the publisher's own claims, not independently verified.
- Matching Gripette hardware on the target arm is stated as required; performance across other arms is not shown.
- Based on bounded source text, not independent testing. Practical suggestions are editorial proposals.