- Generated by
1.16.1
|
Mujoco KDL Wrapper
0.5.1
MuJoCo + KDL bridge for robot kinematics and dynamics
|
The project no longer fetches anything but MuJoCo:
| 0.4 | 0.5 |
|---|---|
| CMake cloned the Orocos KDL fork (MJ_KDL_FETCH_OROCOS_KDL) | run vcs import < mj_kdl_wrapper.repos before configuring; the fork tracks branch vereshchagin-driver-weighting |
| MJ_KDL_FETCH_MENAGERIE, mj-kdl-fetch-menagerie, mj-kdl-fetch-examples | removed; the bundled assets/ are the only models shipped. Point RobotSpec.path at any other MJCF, e.g. a MuJoCo Menagerie checkout |
| Python mjk.menagerie.model_path() / asset_path() and their MJ_KDL_* env overrides | mjk.ASSETS_DIR / "kinova_gen3/gen3.xml" |
| The asset cache was filled by the Menagerie fetch | cmake --install copies assets/ into ~/.cache/mj_kdl_wrapper/assets for programs that do not know the source tree |
| F/T wrench was MuJoCo's reading (the parent's force on the child) | the load on the sensor, as a physical F/T reports it: the negation |
The Env now owns everything the loop touches: the model/data, the robots, the scene slots (env.scene) and the viewer (env.viewer). The loop and its calls take the Env:
| 0.3.x | 0.4 |
|---|---|
| init_robot_from_mjcf(&r, model, data, ...) + env_add_robot(&env, &r) | init_robot_from_mjcf(&r, &env, ...) (registers it) |
| init_window_sim(&viewer, &robot) / (&viewer, m, d) | open_viewer(&env, title) |
| step(&robot), step_n(), step(&viewer, m, d) | step(&env) |
| update(&robot), read_measurements(), apply_commands() | update(&env) (every robot and scene slot) |
| pace_realtime(&robot) / (&viewer, m) | pace_realtime(&env) |
| get_body_frame(m, d, ...), get_site_frame, set_body_pose(m, d, ...) | same names, taking &env |
| init_scene_state(), read/apply/rebind_scene_state() | bind_scene_*(&env.scene, ...); update(&env) and rebuilds handle the rest |
| scene_add_object(&m, &d, &spec, ...) | scene_add_object(&env, ...) |
| record_frame(&vr, m, d), render_rgb(&vr, m, d, out) | record_frame(&vr, &env), render_rgb(&vr, &env, out) |
| set_joint_pos(&r, q, call_forward), mark_kinematics_*() | set_joint_pos(&r, q); frames notice state changes themselves |
| init_window(), render(), cleanup(&viewer) | removed; cleanup(&env) closes the viewer |
reset(&env) now re-seeds every robot's ports and F/T readings and every scene slot, then runs on_reset, and the Simulate UI's reset button does the same. A Robot's MuJoCo index maps are private.
Also changed in 0.4:
| Change | What to do |
|---|---|
| Calls that can fail return mj_kdl::Status instead of bool | if (!s) still works; print s.error |
| Spec string fields (path, prefix, mjcf_path, tool_body, ...) are std::string | drop .c_str(); empty means not set |
| set_body_pose() takes [x, y, z, w] (was [w, x, y, z]) | reorder the quaternion |
| Control modes are MuJoCo actuator groups (RobotSpec::modes, set_control_mode()) | nothing is written to qfrc_applied any more |
| Unset required numbers (timestep, primitive size/mass/friction/rgba, camera pos/fovy) fail the build | set them |
| Earlier docs used tool_body = "g_base" for the 2F-85, which leaves the mount's mass out of KDL | use "g_base_mount" |
| Python: joint ports read as read-only numpy arrays | assign whole vectors; port[i] = x raises |
| Python: Env.save_xml() -> save_model_xml(), Robot.tip_to_tcp -> tip_T_tcp | rename; there are no aliases |
| get_camera_names() removed | loop i < model->ncam with mj_id2name(model, mjOBJ_CAMERA, i) |
| get_joint_position() / get_joint_velocity() removed | a bind_scene_joint() slot's position/velocity, or model/data directly |
| realtime_factor_of() removed | read Viewer::realtime_factor |
| use_camera(VideoRecorder*) / set_free_camera(VideoRecorder*) removed (the Viewer overloads stay) | write vr.cam: type = mjCAMERA_FIXED, fixedcamid = mj_name2id(model, mjOBJ_CAMERA, name); or mjv_defaultFreeCamera() |
| scene_object_site_name() removed (C++ and Python) | the asset's own site name; objects are no longer auto-prefixed: set SceneObject::prefix (Python SceneObject.prefix) to prefix them |
| attach_to_spec(), add_floor_to_spec(), add_skybox_to_spec(), add_objects_to_spec(), compile_and_make_data(), ensure_plugins_loaded() are internal | go through SceneSpec and build_scene(); to run on your own pair, set Env::adopt |
| Python: Env.time(), timestep(), camera_names(), save_binary(), mjk.mujoco_version() removed | env.data.time, env.model.opt.timestep, [env.model.camera(i).name for i in range(env.model.ncam)], mujoco.mj_saveModel(env.model, path, None), mjk.__mujoco_version__ |
| Python: Env.actuator_ctrl(), set_actuator_ctrl(), has_actuator(), set_body_wrench() removed | env.data.actuator(n).ctrl[0] (read or assign), env.model.actuator(n) (KeyError when missing), env.data.body(n).xfrc_applied[:] = [*f, *t] |
| Python: Robot.fk_frame(), gravity_torques() removed | PyKDL ChainFkSolverPos_recursive / ChainDynParam on robot.kdl_chain() with a JntArray of robot.jnt_pos_msr |
| Python: Robot.ft_sensor_frame(name) removed | env.site_frame(<the sensor's frame_site>) |
| find_ft_sensor() removed | search robot.ft_sensors (in ToolFrameSpec::ft_sensors order) by .name |
| MJCF SceneObjects are no longer prefixed with name + "_" | use the asset's own element names; set SceneObject::prefix when an asset is used twice |
| LogLevel is a threshold, INFO < WARN < ERROR < NONE, default INFO (it was a verbosity where ERROR printed everything) | set_log_level(WARN) now prints warnings and errors; the default prints what it did before |
| LOG_INFO() / LOG_WARN() / LOG_ERROR() | MJ_LOG_INFO() / MJ_LOG_WARN() / MJ_LOG_ERROR() |
| The header no longer includes <GLFW/glfw3.h> | include it yourself for the GLFW_KEY_* codes of key_pressed() / capture_key() |
| init_env() dropped an on_reset set before it | it keeps it, like adopt |
| scene_add_object() / scene_remove_object() started from a fresh mjData at qpos0 | the time, qpos/qvel (by joint name and type) and act/ctrl (by actuator name) carry over into the new model; call reset(&env) for the old behaviour |
| A rebuild whose robot or F/T sensor rebinding failed left the Env half rebound | it is left unchanged and the call returns the error |
| scene_add_object() / scene_remove_object() could be called from on_reset | they fail there; change the scene outside the hook |
| A VideoRecorder kept rendering the old model after a rebuild | it builds its render context at the first frame and follows rebuilds |
| init_offscreen() / init_video_recorder() on an open recorder leaked it | they free it first |
| open_viewer() without a usable display could abort on the render thread | it returns an error Status when no window can be created |
| Simulate UI: a reset was detected from time jumping back (also on API reset, rebuild, history scrub, the Load key) | only an explicit UI Reset runs the reset path and on_reset; scrubbing and the Load key no longer reset |
| Simulate UI: the right arrow single-stepped MuJoCo behind the wrapper; a dropped model file was loaded | the right arrow no longer single-steps; drag-and-drop loading is gone (build scenes with SceneSpec) |
| CMake: find_package(mj_kdl_wrapper 0.3) accepted any 0.x (SameMajorVersion) | the minor version must match (SameMinorVersion); ask for 0.4 |
| Install: licenses went to lib/, image_io.hpp and camera_ros.hpp were always installed | licenses in share/doc/mj_kdl_wrapper; only the headers of what was built |
| New: pkg-config file mj_kdl_wrapper.pc | pkg-config --cflags --libs mj_kdl_wrapper; see the Standalone Installation Guide |
| Python: Scene removed | Env.build(spec); its frame, pose, object and save calls are on Env |
| Python: Robot.from_scene(scene, ...) | env.create_robot(...) or env.create_robot_from_chain(...) |
| Python: Robot.update(), step(), pace(), step_n(), Scene.step(), step_n() | env.update(), env.step(), env.pace(); loop for several steps |
| Python: SimulateViewer.open(robot_or_scene, title), .close(), .step(), .pace(), .step_n() | env.open_viewer(title); env.viewer has is_running() and the trace, camera and key calls; env.close() closes it |
| Python: Robot.set_joint_pos(q, call_forward) | set_joint_pos(q) |
| Python: VideoRecorder.open(scene, ...) / open_preset(scene, ...) | pass the Env: open(env, ...) |
| Python: Env.spec was a writable copy whose changes did nothing | it is read-only; change the scene with add_object() / remove_object() |
| Python: Env.on_reset got a ResetContext that dangled once the call returned | it gets a copy, safe to keep |
| Python: menagerie.model_path() knew only kinova_gen3 and robotiq_2f85 | also universal_robots_ur5e and universal_robots_ur10e |
step() no longer sleeps. Pacing moved out of the physics call and into pace_realtime(), so a loop that owns its own timing is no longer fought by a hidden sleep inside step(). A windowed run that should track wall time needs one pace_realtime(&env) call per iteration; headless runs are unaffected, because pacing only ever happened when a viewer existed. See Loop Pacing and the Real-Time Factor.
0.3.0 removes the euler placement field from RobotSpec, AttachmentSpec, SceneObject and CameraSpec, and from attach_child(). Placement orientation is now the quaternion quat, in [x, y, z, w] order, with identity { 0, 0, 0, 1 }. There is no compatibility shim: code setting euler no longer compiles.
quat is the library's own convention. MuJoCo's [w, x, y, z] ordering stays inside the wrapper and is never exposed through these specs.
C++, a 180-degree flip about x:
Python, the same rotation:
For angles without an exact quaternion, convert with SciPy – Rotation.from_euler("xyz", [rx, ry, rz], degrees=True).as_quat() returns [x, y, z, w] directly.
1.16.1