Releases: NVlabs/curobo
Releases · NVlabs/curobo
Isaac Sim 4.0 support, CUDA Graph improvements, minor bug fixes
New Features
- Interpolated metrics calculation now recreates cuda graph if interpolation steps exceed existing buffer size.
- Add experimental
CUDAGraph.reset
usage ascuda>=12.0
is not crashing when an existing captured CUDAGraph is freed and recaptured with new memory pointers. Try this experimental feature by
setting an environment variableexport CUROBO_TORCH_CUDA_GRAPH_RESET=1
. This feature will allow for changing the problem type inmotion_gen
andik_solver
without requiring recreation of the class. - Add partial support for Windows.
- Add Isaac Sim 4.0.0 docker support.
- Examples now work with Isaac Sim 4.0.0.
- Add XRDF support.
- Add curobo.types.file_path.ContentPath to store paths for files representing robot and world. This
improves development on top of cuRobo with custom robots living external of cuRobo library. - Add attach external objects to robot link API to CudaRobotModel.
- Add MotionGenStatus.DT_EXCEPTION to report failures due to trajectory exceeding user specified
maximum trajectory dt. - Add reading of end-effector mesh if available when rendering trajectory with
UsdHelper
, also
supports goalset rendering. - Kinematics module (
curobo.cuda_robot_model
) has complete API documentation.
Changes in Default Behavior
- Cuda graph capture of optimization iterations resets solver before recording.
join_path(a, b)
now requiresa
to not have a trailing slash to make the file compatible with Windows.- Drop examples support for Isaac Sim < 4.0.0.
- asset_root_path can be either empty string or None.
- Order of variables in
SelfCollisionKinematicsConfig
has changed. Unused variables
moved to bottom. - Remove requirement of warmup for using
offset_waypoint
inPoseCost
.
BugFixes & Misc.
- Minor documentation fixes to install instructions.
- Add support for older warp versions (<1.0.0) as it's not possible to run older isaac sim with newer warp versions.
- Add override option to mpc dataclass.
- Fix bug in
PoseCost.forward_pose()
which causedtorch_layers_example.py
to fail. - Add warp constants to make module hash depend on robot dof, for modules that generate runtime
warp kernels. This fixes issues using cuRobo in isaac sim. - Add
plan_config.timeout
check toplan_single_js()
. - Recreation of interpolation buffer now copies the joint names from raw trajectory.
- Fix bug in running captured cuda graph on deleted memory pointers
when getting metrics on interpolated trajectory - Change order of operations in cuda graph capture of particle opt to get correct results
during graph capture phase. - Franka Panda now works in Isaac Sim 4.0.0. The fix was to add inertial parameters to all links in
the urdf. - Create new instances of rollouts in wrap classes to ensure cuda graph rollouts are not
accidentally used in other pipelines. - Add cuda graph check for
get_metrics
. - Remove aligned address assumption for float arrays inside kernel (local memory).
- Add check for existing warp kernel in a module before creating a new one to avoid corruption of
existing cuda graphs.
Improved joint space planning and planning start validity check
New Features
- Add start state checks for world collision, self-collision, and joint limits.
- Add finetune with dt scaling for
motion_gen.plan_single_js
to get more time optimal
trajectories in joint space planning. - Improve joint space planning convergence, now succeeds in more planning problems with higher
accuracy.
Changes in default behavior
- Some warp kernels are now compiled based on runtime parameters (dof), causing a slowdown in load
time for motion_gen. To avoid this slowdown, add an environment variableCUROBO_USE_LRU_CACHE=1
which will cache the runtime generated kernels.
BugFixes & Misc.
- Fix bug in evaluator to account for dof maximum acceleration and jerk.
- Add unit test for different acceleration and jerk limits.
- Add a check in self-collision kernels to avoid computing over inactive threads.
- Add
link_poses
as an additional property to kinematics to be more descriptive. - Add
g_dim
check forint
in batched planning. - Add
link_poses
for motion_gen.warmup() in batch planning mode. - Add
link_poses
as input tobatch_goalset
. - Add finetune js trajopt solver.
- Pass raw velocity, acceleration, and jerk values to dt computation function to prevent
interpolation errors from causing out of joint limit failures - Add
finetune_js_dt_scale
with a default value > 1.0 as joint space trajectories are
time optimal in sparse obstacle environments. - Add note on deterministic behavior to website. Use lbfgs history < 12 for deterministic
optimization results. - Add warning when adding a mesh with the same name as in existing cache.
- Remove warmup for batch motion gen reacher isaac sim example.
- Fix python examples in getting started webpage.
- Refactor warp mesh query kernels to use a
wp.func
for signed distance queries.
Re-timing post-processing improves motion generation quality at slow speeds, API documentation
New Features
- Significant improvements for generating slow trajectories. Added re-timing post processing to
slow down optimized trajectories. UseMotionGenPlanConfig.time_dilation_factor<1.0
to slow down a
planned trajectory. This is more robust than settingvelocity_scale<1.0
and also allows for
changing the speed of trajectories between planning calls curobo.util.logger
addslogger_name
as an input, enabling use of logging api with other
packages.
Changes in default behavior
- Move
CudaRobotModelState
fromcurobo.cuda_robot_model.types
to
curobo.cuda_robot_model.cuda_robot_model
- Activation distance for bound cost in now a ratio instead of absolute value to account for very
small range of joint limits whenvelocity_scale<0.1
. TrajResult
is renamed toTrajOptResult
to be consistent with other solvers.- Order of inputs to
get_batch_interpolated_trajectory
has changed. MpcSolverConfig.load_from_robot_config
usesworld_model
instead ofworld_cfg
to be
consistent with other wrappers.
BugFixes & Misc.
- Fix bug in
MotionGen.plan_batch_env
where graph planner was being set to True. This also fixes
isaac sim examplebatch_motion_gen_reacher.py
. - Add
min_dt
as a parameter toMotionGenConfig
andTrajOptSolverConfig
to improve readability
and allow for having smallerinterpolation_dt
. - Add
epsilon
tomin_dt
to make sure after time scaling, joint temporal values are not exactly
at their limits. - Remove 0.02 offset for
max_joint_vel
andmax_joint_acc
inTrajOptSolver
- Bound cost now scales the cost by
1/limit_range**2
whenlimit_range<1.0
to be robust to small
joint limits. - Added documentation for
curobo.util.logger
,curobo.wrap.reacher.motion_gen
,
curobo.wrap.reacher.mpc
, andcurobo.wrap.reacher.trajopt
. - When interpolation buffer is smaller than required, a new buffer is created with a warning
instead of raising an exception. torch.cuda.synchronize()
now only synchronizes specified cuda device with
torch.cuda.synchronize(device=self.tensor_args.device)
- Added python example for MPC.
Improved motion quality, precision, and better joint space planning.
New Features
- Add mimic joint parsing and optimization support. Check
ur5e_robotiq_2f_140.yml
. - Add
finetune_dt_scale
as a parameter toMotionGenPlanConfig
to dynamically change the
time-optimal scaling on a per problem instance. MotionGen.plan_single()
will now try finetuning in a for-loop, with larger and larger dt
until convergence. This also warm starts from previous failure.- Add
high_precision
mode toMotionGenConfig
to support<1mm
convergence.
Changes in default behavior
- collision_sphere_buffer now supports having offset per link. Also, collision_sphere_buffer only
applies to world collision while self_collision_buffer applies for self collision. Previously,
self_collision_buffer was added on top of collision_sphere_buffer. TrajEvaluatorConfig
cannot be initialized without dof as now per-joint jerk and acceleration
limits are used. UseTrajEvaluatorConfig.from_basic()
to initialize similar to previous behavior.finetune_dt_scale
default value is 0.9 from 0.95.
BugFixes & Misc.
- Fix bug in
WorldVoxelCollision
whereenv_query_idx
was being overwritten. - Fix bug in
WorldVoxelCollision
where parent collision types were not getting called in some cases. - Change voxelization dimensions to include 1 extra voxel per dim.
- Added
seed
parameter toIKSolverConfig
. - Added
sampler_seed
parameterRolloutConfig
. - Fixed bug in
links_goal_pose
where tensor could be non contiguous. - Improved
ik_solver
success by removing gaussian projection of seed samples. - Added flag to sample from ik seeder instead of
rollout_fn
sampler. - Added ik startup profiler to
benchmark/curobo_python_profile.py
. - Reduced branching in Kinematics kernels and added mimic joint computations.
- Add init_cache to WorldVoxelCollision to create cache for Mesh and Cuboid obstacles.
TrajEvaluator
now uses per-joint acceleration and jerk limits.- Fixed regression in
batch_motion_gen_reacher.py
example where robot's position was not being
set correctly. - Switched from smooth l2 to l2 for BoundCost as that gives better convergence.
requires_grad
is explicitly stored in a varaible beforetensor.detach()
in warp kernel calls
as this can get set to False in some instances.- Fix dt update in
MotionGen.plan_single_js()
where dt was not reset after finetunestep, causing
joint space planner to fail often. - Improve joint space planner success by changing smooth l2 distance cost to l2 distance. Also,
added fallback to graph planner when linear path is not possible. - Retuned weigths for IKSolver, now 98th percentile accuracy is 10 micrometers wtih 16 seeds
(vs 24 seeds previously). - Switch float8 precision check from
const
to macro to avoid compile errors in older nvcc, this
fixes docker build issues for isaac sim 2023.1.0.
v0.7.0 with improved collision gradients, ESDF Voxel grid, Constrained Planning, Isaac sim 2023.1.1 support
New Features
- Add function to disable and enable collision for specific links in KinematicsTensorConfig.
- Add goal index to reacher results to return index of goal reached when goalset planning.
- Add locked joint state update api in MotionGen class.
- Add goalset warmup padding to handle varied number of goals during goalset planning and also when
calling plan_single after warmup of goalset. - Add new trajopt config to allow for smooth solutions at slow speeds (
velocity_scale<=0.25
). Also
add error whenvelocity_scale<0.1
. - Add experimental robot image segmentation module to enable robot removal in depth images.
- Add constrained planning mode to motion_gen.
- Use
torch.compile
to leverage better kernel fusion in place oftorch.jit.script
. - Significantly improved collision computation for cuboids and meshes. Mesh collision checker is
now only 2x slower than cuboid (from 5x slower). Optimization convergence is also improved. - LBFGS kernels now support
history <= 31
fromhistory <= 15
. - 2x faster LBFGS kernel that allocates upto 68kb of shared memory, preventing use in CUDA devices
with compute capability<7.0
. - On benchmarking Dataset, Planning time is now 42ms on average from 50ms. Higher quality solutions
are also obtained. See benchmarks for more details. - Add
WorldCollisionVoxel
, a new collision checking implementation that uses a voxel grid
of signed distances (SDF) to compute collision avoidance metrics. Documentation coming soon, see
benchmark/curobo_voxel_benchmark.py
for an example. - Add API for ESDF computation from world representations, see
WorldCollision.get_esdf_in_bounding_box()
. - Add partial support for isaac sim 2023.1.1. Most examples run for UR robots.
Franka Panda
is
unstable.
Changes in default behavior
- Increased default collision cache to 50 in RobotWorld.
- Changed
CSpaceConfig.position_limit_clip
default to 0 as previous default of 0.01 can make
default start state in examples be out of bounds. - MotionGen uses parallel_finetune by default. To get previous motion gen behavior, pass
warmup(parallel_finetune=False)
andMotionGenPlanConfig(parallel_finetune=False)
. - MotionGen loads Mesh Collision checker instead of Primitive by default.
- UR10e and UR5e now don't have a collision sphere at tool frame for world collision checking. This
sphere is only active for self collision avoidance. - With torch>=2.0, cuRobo will use
torch.compile
instead oftorch.jit.script
to generate fused
kernels. This can take several seconds during the first run. To enable this feature, set
environment variableexport CUROBO_TORCH_COMPILE_DISABLE=0
.
Breaking Changes
- Renamed
copy_if_not_none
toclone_if_not_none
to be more descriptive. Nowcopy_if_not_none
will try to copy data into reference. - Renamed
n_envs
in curobo.opt module to avoid confusion between parallel environments and
parallel problems in optimization. - Added more inputs to pose distance kernels. Check
curobolib/geom.py
. - Pose cost
run_vec_weight
should now be[0,0,0,0,0,0]
instead of[1,1,1,1,1,1]
max_distance
is now tensor fromfloat
and is an input to collision kernels.- Order of inputs to
SweptSdfMeshWarpPy
has changed.
BugFixes & Misc.
- refactored wp.index() instances to
[]
to avoid errors in future releases of warp. - Fix bug in gaussian transformation to ensure values are not -1 or +1.
- Fix bug in ik_solver loading ee_link_name from argument.
- Fix bug in batch_goalset planning, where pose cost was selected as GOALSET instead of
BATCH_GOALSET. - Added package data to also export
.so
files. - Fixed bug in transforming link visual mesh offset when reading from urdf.
- Fixed bug in MotionGenPlanConfig.clone() that didn't clone the state of parallel_finetune.
- Increased weighting from 1.0 to 10.0 for optimized_dt in TrajEvaluator to select shorter
trajectories. - Improved determinism by setting global seed for random in
graph_nx.py
. - Added option to clear obstacles in WorldPrimitiveCollision.
- Raise error when reference of tensors change in MotionGen, IKSolver, and TrajOpt when cuda graph
is enabled. - plan_single will get converted to plan_goalset when a plan_goalset was used to initialize cuda
graph. - plan_goalset will pad for extra goals when called with less number of goal than initial creation.
- Improved API documentation for Optimizer class.
- Set
use_cuda_graph
toTrue
as default fromNone
inMotionGenConfig.load_from_robot_config
- Add batched mode to robot image segmentation, supports single robot multiple camera and batch
robot batch camera. - Add
log_warn
import toarm_reacher.py
- Remove negative radius check in self collision kernel to allow for self collision checking with
spheres of negative radius. - Added
conftest.py
to disabletorch.compile
for tests. - Added UR5e robot with robotiq gripper (2f-140) with improved sphere model.
- Fix bug in aarch64.dockerfile where curobo was cloned to wrong path.
- Fix bug in aarch64.dockerfile where python was used instead of python3.
- Remove unused variables in kernels.
- Added
pybind11
as a dependency as some pytorch dockers for Jetson do not have this installed. - Fix incorrect dimensions in
MotionGenResult.success
inMotionGen.plan_batch()
when
trajectory optimization fails. - Added unit tests for collision checking functions.
- Fix bug in linear interpolation which was not reading the new
optimized_dt
to interpolate
velocity, acceleration, and jerk. - Remove torch.jit.script wrapper for lbfgs as it causes TorchScript error if history is different
between trajopt and finetune_trajopt.
Known Bugs (WIP)
Franka Panda
robot loading from urdf in isaac sim 2023.1.1 is unstable.
New release with bug fixes, features, & higher quality solutions
New Features
- Added support for actuated axis to be negative (i.e., urdf joints with
<axis xyz="0 -1 0"/>
are
now natively supported). - Improved gradient calculation to account for terminal state. Trajectory optimization can reach
within 1mm of accuracy (median across 2600 problems at 0.017mm). - Improved estimation of previous positions based on start velocity and acceleration. This enables
Trajectory optimization to optimize from non-zero start velocity and accelerations. - Added graph planner and finetuning step to joint space planning (motion_gen.plan_single_js). This
improves success and motion quality when planning to reach joint space targets. - Added finetuning across many seeds in motion_gen, improving success rate and motion quality.
- Add urdf support to usd helper to export optimization steps as animated usd files for debugging
motion generation. Checkexamples/usd_examples.py
for an example. - Retuned weights for IK and Trajectory optimization. This (+ other fixes) significantly improves
pose reaching accuracy, IK accuracy improves by 100x (98th percentile < 10 micrometers) and motion
generation median at 0.017mm (with). IK now solves most problems with 24 seeds (vs 30 seeds prev.).
Runbenchmark/ik_benchmark.py
to get the latest results. - Added
external_asset_path
to robot configuration to help in loading urdf and meshes from an
external directory.
BugFixes & Misc.
- Update nvblox wrappers to work with v0.0.5 without segfaults. Significantly improves stability.
- Remove mimic joints in franka panda to maintain compatibility with Isaac Sim 2023.1.0 and 2022.2.1
- Cleanup docker scripts. Use
build_docker.sh
instead ofbuild_dev_docker.sh
. Added isaac sim
development docker. - Fixed bug in backward kinematics kernel, helped improve IK and TO pose reaching accuracy..
- Changed
panda_finger_joint2
from<axis xyz="0 1 0"/>
to<axis xyz="0 -1 0"/>
infranka_panda.urdf
to match real robot urdf as cuRobo now supports
negative axis. - Changed benchmarking scripts to use lock joint state of [0.025,0.025] for mpinets dataset.
- Added scaling of mesh to Mesh.get_trimesh_mesh() to help in debugging mesh world.
- Improved stability and accuracy of MPPI for MPC.
- Added NaN checking in STOMP covariance computation to account for cases when cholesky decomp
fails. - Added ground truth collision check validation in
benchmarks/curobo_nvblox_benchmark.py
.
Performance Regressions
- cuRobo now generates significantly shorter paths then previous version. E.g., cuRobo obtains
2.2 seconds 98th percentile motion time on the 2600 problems (benchmark/curobo_benchmark.py
), where
previously it was at 3 seconds (1.36x quicker motions). This was obtained by retuning the weights and
slight reformulations of trajectory optimization. These changes have led to a slight degrade in
planning time, 20ms slower on 4090 and 40ms on ORIN MAXN. We will address this slow down in a later
release. One way to avoid this regression is to setfinetune_dt_scale=1.05
in
MotionGenConfig.load_from_robot_config()
.
Isaac Sim 2023.1.0 and dockerfiles
- Added support for Isaac Sim 2023.1.0 in examples
- Added dockerfiles for python and isaac sim installation
Initial Public Release
v0.6.0 Update issue templates