Grasp, move and coordinate on cycle time

Getting a grip on real parts, moving without collisions, and keeping two arms out of each other’s way, within the speed and force limits a shared workspace needs.

Start a project How we choose models

Maximum grip force of a Robotiq 2F-85. Most failures to hold a light part are friction or geometry, not force.: 235 NMaximum grip force of a Robotiq 2F-85. Most failures to hold a light part are friction or geometry, not force.
The licence of KDL, MoveIt’s default inverse-kinematics plugin. Almost nobody notices; we check.: LGPL-2.1The licence of KDL, MoveIt’s default inverse-kinematics plugin. Almost nobody notices; we check.
Speed-and-separation and power-and-force limits we design to when people and arms share a space.: ISO/TS 15066Speed-and-separation and power-and-force limits we design to when people and arms share a space.

What we do

Grasping solutions
Reliable grasps on the parts you actually have, including dense and messy bins.
- Parallel-jaw, suction and multi-finger grippers
- Grasp planning in dense, cluttered bins
- Pushing and singulation before the grasp
- Collision and reachability checks on every grasp
Motion planning and arm control
Collision-free paths and controllers that handle contact.
- MoveIt 2 with OMPL and Pilz planners
- GPU motion generation with cuRobo
- Time-optimal trajectories with Ruckig and TOPP-RA
- Force, impedance and compliance control
Multi-arm movement and coordination
Two or more arms sharing a workspace safely.
- Shared-workspace collision avoidance
- Synchronized trajectories and handovers
- Bimanual policies for two-handed tasks

Model or no model

Where should the gripper close, and how does the arm get there?

Programmed approach

Grasps precomputed from CAD or simple shapes
MoveIt 2 planners and inverse kinematics
Trajectory optimization for cycle time

Learned approach

Grasp proposals predicted from point clouds
Learned grasp-success scoring
Policies for contact-rich final motions

How we choose

Known parts in known poses get precomputed grasps, and free-space moves stay programmed because they are fast and verifiable. Mixed objects in dense bins get learned grasp proposals, each filtered by collision and reachability checks before the arm moves.

Grasp models and their catches

A grasp model is trained for a particular gripper, and its licence decides whether you can ship it.

Model	Gripper it assumes	Catch
Contact-GraspNet	Franka hand	Re-check grasps for your jaw width
GraspGen	Gripper given as an input	Needs a GPU
AnyGrasp	Parallel jaw	Strong results, but a machine-locked licence key
GPD	Parallel jaw	BSD licence, runs without a GPU, older method