Dev - #155
Open
vikashplus wants to merge 76 commits into
Open
Conversation
Point cloud utility functions for mujoco [Draft]
… compliance. Envs now overwrite the correct _reset function. Gym compliance is now preserved across the defaulting envs
- gym/gymnasium compatibility doesn't port over when custom reset is provided - we are now recommending a call to super().reset(**kwargs) to get the appropritae return type ensuring compatibility
np.product -> np.prod for np 2.0 compatibility
bump pyh5 and pin numpy
1. FEATURE: Explicitely marks the closure of the trace
2. BUGFIX: Introduces create_datum. create_dataset is updated and is now reserved for adding full datasets (shouldn't be edited later). If appending is needed first create_datum() and then follow it with repeated calls to append_datum().
3. BUGFIX: verify_type is stricter now. It goes beyond checking for root dtypes. It thoroughly checks for type, shapes, keys - even for nested structures. This implies that dict{datum} also needs to maintain its internal keys to support append leading to further consistency
4. FEATURE: Much like rendering of rgb datasets, now trace supports plotting of numerical datasets using a list(grp_keys)_ng and list(list(dataset_keys))_ng
5. Adds multiple tests to check for data consistency and plotting consistency
Collaborator
|
… can be further improved by having a global settings of rnumerical precision throughout the repo
Minor upgrades
dependency update and code alignment
…eck for existance of 'time' key for sensors
…rry complex data from the hardware into sensordata field of the model. Reply on sensordata field to create obs and obs_dicts
…e 0-d scalars squeeze_dims() previously did a blanket np.squeeze(), which removed every singleton dimension instead of just the (num_traj=1, horizon=1) wrapper added by expand_dims(). Single-value observations (e.g. shape (1,)) were squeezed all the way down to 0-d scalars, which then failed to concatenate with other proprio/obs vectors in get_proprioception() and obsdict2obsvec(). Now only squeeze the leading axes that expand_dims actually added, leaving real data dimensions (and pre-existing scalar rwd_dict entries) untouched.
Robot._resolve_data_ids caches each joint actuator's feedback address as data_id into sim.data.qpos, but computed it with jnt_dofadr (the qvel address). qpos and qvel addresses coincide only while all preceding joints are 1-DOF: any free/ball joint earlier in the model (e.g. a manipulable object defined before the arm) shifts every subsequent actuator's read by the qpos/qvel size difference (freejoint: 7 vs 6). The stale read feeds the velocity limiter in process_actuator(), which re-anchors each command to the wrong joint's position — in practice the arm freezes at whatever pose keeps the misread error small. Observed on anthrohive URDtriosPickPlace-v0 (box freejoint precedes the UR5e): the arm ignored all position targets until this fix.
Extend supported sensors
fix(robot): index actuator qpos reads with jnt_qposadr, not jnt_dofadr
…e base class. So there is no need to deviate away from the get_sensor() signature
…handling of outputs. Better plots
- robohive/robot/hardware_base.py — new HARDWARE_REGISTRY/SENSOR_POSTPROCESS dicts and @register_hardware(type_name, sensor_postprocess=None) decorator (works on classes or factory functions). - robohive/robot/hardware_franka.py, hardware_robotiq.py, hardware_optitrack.py, hardware_realsense.py, hardware_realsense_single.py, hardware_dynamixel.py — each now conforms to hardwareBase (adds recover(), fixes get_sensors() naming/keys to the neutral pos/vel/effort contract) and self-registers via @register_hardware(...). - hardware_optitrack.py also had a pre-existing broken import (darwin.darwin_robot...) fixed to robohive.robot.hardware_base. - hardware_dynamixel.py's Dynamixels was previously non-functional. Added a real adapter. It will need a consolidation effort to reconsile all dynamixel based robots
- hardware_init, hardware_get_sensors, hardware_apply_controls, hardware_close collapsed from hardcoded if/elif chains on interface['type'] into uniform registry-driven loops. - New Robot.hardware_reset(reset_pos): a dedicated blocking/large-displacement reset path (each hardware class runs its own min-jerk trajectory), replacing the old hardware_apply_controls(..., is_reset=True) overload. - Fixes a real bug where reset indexed by hdr_id (counts all actuators) while the ctrl vector it built only counted qpos-type actuators — silently misaligned on any device mixing qpos and tendon actuators (e.g. arm+gripper). - get_visual_sensors() now duck-types hasattr(device['robot'], 'get_frame') instead of asserting interface['type'] == 'realsense', so UVC/webcam-style cameras (rgb-only, no depth) no longer crash it. - adr renamed to hdr_adr throughout (configure_robot(), camera calibration) — now consistently documented as a physical hardware address (motor ID, RTDE join- adr renamed to hdr_adr throughout (confe raise NotImplemented → raise NotImplementedError(...) with descriptive messages.
- Added robot_cls = kwargs.pop('robot_cls', Robot) extension point so downstream repos never need to subclass/swap Robot.
- Added optional init_qpos param — if provided, used directly as self.init_qpos; otherwise falls back to today's auto-computed mid-actuator-range behavior. Warns if hardware-backed with no explicit init_qpos (homing to an implicit pose is likely unintended on real hardware).
- Swapped the startup call from self.step(np.zeros(...)) to self.reset() + self.forward(), so hardware now goes through Robot.reset()'s min-jerk hardware_reset() on env construction instead of jumping straight to a raw zero-ctrl command from wherever the robot currently is.
… internally (to route hardware envs through min-jerk reset) instead of self.step(zeros). Since reset() is often overridden by subclasses, this exposed a latent ordering bug across ~20 env files: they computed a task-specific self.init_qpos after calling super()._setup(), but by then super()._setup() had already triggered the first reset() using a stale/default init_qpos. In relocate_v0.py this stale pose left the hand and object permanently in collision, and the env's collision-retry logic in reset() recursed forever (RecursionError).
Vk/consolidate robot - new hardware registery - hardware classes follow base class contracts
FEATURE: Improved path_utils. Better scanning of directories. Better …
…ell as calibration hardware_reset now takes an explicit space='sim'|'hdr' parameter and indexes reset_pos via data_id (sim) or hdr_id (hdr), matching the pattern already used in hardware_apply_controls. It now applies actuator['scale']/actuator['offset'] before handing the position to device['robot'].reset(...), same conversion hardware_apply_controls does for per-step controls.
…like a copy-paste typo. Fixed
BUGFIX robots (hardware-reset-space-&-calib-bug +vel_mode-typo)
…ng hardware connections - env_base.py: close() now also closes robot (hardware) and sim_obsd (when it differs from sim), instead of only closing sim. Added close_hardware=True arg so callers can leave a hardware connection open across close()/make() cycles intentionally (env.close(close_hardware=False)). Added a warn-only __del__, mirroring Robot's, so a hardware-backed env that's GC'd without close() ever being called surfaces a RuntimeWarning instead of silently dangling. - robot.py: Robot.__init__ was assigning the persistent session via self.robot_config, which only ever shadowed the class attribute on that one instance and silently defeated reuse across separate Robot()/env.make() calls. Now assigns via type(self).robot_config so the session is genuinely shared. Robot.close() gained the matching close_hardware arg, and tracks _explicitly_closed so __del__ doesn't false-positive when hardware was left
…uld get silently swallowed
- renderer.py/mj_renderer.py: replaced the private MJRenderer._user_exit flag with
a base-class Renderer.exit_requested property so any backend can surface it.
Also set it when the viewer window's native close ('X') button is used, not just
Escape, which previously did nothing.
- env_base.py: added MujocoEnv.viewer_exit_requested. examine_policy()/
examine_policy_new() now check it in their rollout while-loop (and again to break
the outer episode loop), instead of continuing to step invisibly after the window
closes. For offscreen/none rendering, which has no window at all, both loops now
catch KeyboardInterrupt and stop the same way, returning whatever was completed.
Offscreen video writes are trimmed to the frames actually rendered before an
early stop instead of writing out the pre-allocated zero-filled tail.
- mj_sim_scene.py: SimScene.advance() had a bare , which also swallows
KeyboardInterrupt/SystemExit, not just real errors. That meant a Ctrl+C landing
inside sim.step() got eaten and turned into a silent physics reset instead of
reaching the new rollout-loop handler. Narrowed to dm_control's PhysicsError and
included the underlying exception in the warning message
…efault Robot class Yesterday's change to persist robot_config on the class (so a hardware session survives an env.close()/make() cycle) also applied to sim-only robots, which all default to the same Robot class. Creating one sim env after another in the same process left the second one silently reusing the first's sensor/actuator config -- e.g. test_arms.py would crash with an IndexError once a larger model's config leaked into a smaller model's sim.data.sensordata. - robot.py: scope the class-level persistence to is_hardware sessions only; sim sessions get a per-instance robot_config that can't leak. close() now also clears whichever level actually holds the config (class attr for hardware, instance attr for sim) instead of only ever touching the class attr. - test_envs.py: check_env/check_old_envs now call env.close() instead of relying on GC via del(), exercising the real teardown path.
Cleaning up exits
BUGFIX: Better handling of x-axis when keys are empty
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
6 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.