Release Notes#

Release 1.1.0 - April 10, 2026#

Added support for NVIDIA DGX Spark.
Added support for Microsoft Windows.
Added wheel files for all actively-supported Python versions (currently 3.10 through 3.14) for all platforms.
Added support for batch mode in collision-free IK.
Added multi-target support for RMPflow.
Added self-collision support for geometric motion planning. Self-collision checking is now the default for both MotionPlanner and geometric planning within the TrajectoryOptimizer pipeline.
Improved performance of nearest-neighbor queries for geometric planning. This optimization is only enabled when using unweighted distance_metric_weights (i.e., a vector of all ones). This is the default, generally-recommended setting for MotionPlanner.
Improved performance of trajectory optimization.
Updated default parameters for collision-free IK and trajectory optimization to improve success rates and convergence.
Extended WorldInspector to support querying obstacle pose() and isEnabled() status.
Enabled Python interface compatibility with NumPy 2.x by updating to pybind11 version 3.0.1.
Eliminated all internal use of CUDA Unified Memory. On Jetson Orin, this enables cuMotion to run in the same process as other threads making simultaneous use of CUDA. Platforms with full Unified Memory support (i.e., x86_64 and Jetson Thor) are unaffected by this change.

The following items are deprecated and will be removed in the next release:

CollisionFreeIkSolver::solveGoalset() and related target constructors have been deprecated in favor of CollisionFreeIkSolver::solveArray() and related target constructors. The newly added solveArray() supports both batch mode (i.e., multiple problems with a single target per problem) and goalset mode (i.e., a single problem with multiple targets). Goalset batch mode support (i.e., multiple problems, each with multiple targets) will be added in a future release.
The single-target interface in RMPflow (RmpFlow::setEndEffectorPositionAttractor(), etc.) has been deprecated in favor of the new multi-target interface (RmpFlow::addTargetFrame(), RmpFlow::setPositionTarget(), etc.).
WorldInspector::distancesTo() now expects distances to be input by reference. A deprecated overload supports pointer input.
The MotionPlannerConfig parameter cuda_tree_params/max_num_nodes has been deprecated and will now be ignored (with warning). The size of CUDA-accelerated trees is no longer directly bounded and cuda_tree_params/max_num_nodes will be ignored.

A CPU memory leak is exposed when adding and removing obstacles to a world. Each cycle adding and removing an obstacle will grow memory usage, with impacts most noticeable for SDF obstacles. Updating values of an existing SDF does not contribute to this memory leak.
Open3D is used for visualization in some of the Python examples. Because Open3D is not currently available from PyPI for the aarch64 architecture, these examples will lack visualization when run on NVIDIA Jetson or NVIDIA DGX Spark.

Improved performance of both collision-free IK and trajectory optimization.
Added experimental support for segmenting (masking) out the contribution of a robot from a depth image. This functionality is currently only exposed in C++ and only supports depth image buffers stored on the host (i.e., in CPU memory).
Eliminated all internal use of the default CUDA stream.

On Jetson Orin, cuMotion should not be run in the same process as other threads making simultaneous use of CUDA. This restriction is related to the manner in which cuMotion uses CUDA Unified Memory and will be lifted in a future release.
Open3D is used for visualization in some of the Python examples. Because Open3D is not currently available from PyPI for the aarch64 architecture, these examples will lack visualization when run on NVIDIA Jetson.

Open3D is used for visualization in some of the Python examples. Because Open3D is not currently available from PyPI for the aarch64 architecture, these examples will lack visualization when run on NVIDIA Jetson.