Project highlights: 28 September to 4 October
- Source: Open Source Robotics
- Published: 2026-10-05
- Category: upstream
PX4 and ArduPilot both changed core naming and addressing, Isaac Lab cut its installer down to uv, and several robots gained simulation models that sit closer to the hardware.
PX4 Autopilot
!PX4 Autopilot
- On the main branch, GPS is now called GNSS in the messages, topics and sensor parameters. The ROS 2 topic is /fmu/out/vehicle_gnss with a new VehicleGnss message, and the translation node cannot bridge the old vehicle_gps_position name, so offboard code that reads the position topic needs updating. The SENS_GPS parameters become SENS_GNSS and are migrated when parameters are imported. - GNSS blending has been removed, along with SENS_GNSS_MASK and SENS_GNSS_TAU. A vehicle with two receivers now uses one at a time, the SENS_GNSS_PRIME receiver while it has a 3D fix and otherwise the one with the best fix, so the position no longer drifts as the blend weights change. - Multicopters get rate-loop gain compression, which lowers the gains when the torque setpoint starts to oscillate. It is off by default and set through the MC_GC_ parameters. PX4 also now warns by default when the remaining flight time runs low, where COM_FLTT_LOW_ACT used to default to no action: Changes.
Isaac Lab
!Isaac Lab
- On the develop branch, uv is now the only way to install Isaac Lab. isaaclab.sh, isaaclab.bat, the --install option and the conda setup are gone, so install with uv sync or uv pip and run everything through uv run isaaclab. - Tasks and tools such as zero_agent no longer open a viewer on their own. Pass --viz with the visualiser you want, for example --viz newton_gl, or nothing is shown. --video now takes a source, so --video viz:kit records the Kit window and --video sensor:wrist_camera records that camera. - NVIDIA Isaac GR00T N1.7 is now supported through RLinf alongside N1.5, each in its own environment. If you already have an RLinf 0.2 setup, run the shared installation step again: Changes.
IR-SIM
!IR-SIM
- Version 2.12.0 adds a contact mode, set with collision_mode: contact. Objects are pushed apart using their mass, friction, inertia and restitution, so a robot can push a box, stall against one that is too heavy and spin one it hits off-centre. Every object reports what it touches through obj.contact. - Lidar scans now carry a per-beam valid mask, and env.get_msg(use_inf=True) reports misses as infinite ranges, as ROS expects under REP 117. Noise is now applied to hits only and every range stays between range_min and range_max, so scans from the same scene may differ from before. - Differential and Ackermann robots under rvo and sfm used to turn away from their goal and now head towards it, and a robot at rest no longer drives into an approaching neighbour. Crowds can be stepped as one group with the social force model, and the default colours have changed to the Okabe-Ito palette: Release notes.
GENESIS
!GENESIS
- Version 1.4.3 makes rigid-body simulation faster on the GPU for a single scene and for batches up to about 4,000 environments, and it uses less GPU memory, which leaves room for larger batches on smaller cards. Compilation on Windows is up to three times faster. - Trajectories and checkpoints recorded with earlier versions no longer load, because the memory work removed arrays they stored. gs.init() also stops setting PyTorch's global default device and dtype, so code that relied on new tensors landing on the Genesis device must pass device=gs.device itself. - Several sensor readings are corrected. Camera extrinsics now update when the camera moves, IMU readings include cross-axis coupling and are recorded properly, and log-scaled depth images are normalised correctly: Release notes.
ArduPilot
!ArduPilot
- The master branch now supports 32-bit MAVLink system IDs, wide enough to use an IPv4 address as the ID, which is aimed at large fleets such as drone light shows. Setting MAV_SYSID above 255 makes every packet unreadable to older MAVLink 2 software, so peripherals that have not been updated will drop them. - Copter's coordinated turns in Loiter mode are now off by default (LOIT_OPTIONS is 0), after mixed feedback from pilots. The developers expect to ship this in 4.7.3, so set the option yourself if you want to keep them. - Plane gets AP_CustomControl, a place to compile in your own controller, pick it with CC_TYPE and turn it on from an auxiliary switch, as Copter already allows. It comes with a working PID example and is not part of standard builds: Changes.
Upkie
!Upkie
- Version 13.0.0 moves the reinforcement-learning code out to a separate project, rl_zoo3_balancer. The upkie.envs.wrappers module and the Stable-Baselines3 dependency are gone, and every environment now returns a reward of zero, so training code has to bring its own reward and wrappers. - The build instructions have moved from the old wiki into the repository's documentation: 3D printing, the bill of materials, cables, electronics testing, Raspberry Pi setup and assembly now sit beside the code they belong to. - The Makefile no longer reads UPKIE_NAME and expects a host called upkie in your SSH config. Version 13.0.1 followed to fix the packaged Bullet simulation spine: Release notes.
AI Worker
!AI Worker
- Version 2.2.8 adds MuJoCo models of the BG2, SG2 and SH5 variants to ffw_description, beside the URDFs. Force limits follow the published motor specifications, while the actuator gains and joint damping are stated to be approximate. - The HX5-D20 hand descriptions now match revision 2 of the hand. Its joints are numbered, not lettered (finger_l_joint_a is now finger_l_joint1), so code that addresses the old names will break. The second thumb joint's upper limit drops from 3.14 to 1.57 rad, and the index and middle finger bases move by 7 mm. - The same day, OpenManipulator-X 5.1.3 added MuJoCo models for the OpenMANIPULATOR-X, OMX and OMY arms, and MuJoCo Menagerie gained the ROBOTIS OMY-3M: Release notes.
wuji-mjlab
!wuji-mjlab
- Version 2026.9.27 adds the Wuji Hand 2. It can turn a cube in the hand (WujiHand2_Reorient, with a 50 Hz variant) and spin a pen (WujiHand2_PenSpin), which comes with 40 motion clips and five sets of pretrained checkpoints. - The release ships ONNX policies and deployment configurations for both hand generations, and a bundle with the hand mounting jigs, cube and pen CAD and 3MF files and a dimensioned assembly drawing. On the real hand, the reorientation script now waits for you to press Enter after homing before the policy takes control. - Training moves to mjlab 1.6.0, MuJoCo 3.11.0 and the official rsl-rl-lib 5.4.2 in place of a vendored copy. It needs Linux, an NVIDIA GPU with CUDA 12.8 and pixi, which is the only supported installer: Release notes.
ORCA Hand
!ORCA Hand
- Version 0.5.0 of orca_core sets the DYNAMIXEL return delay to 20 µs when it configures the motor chain, which cuts a read of all 17 motors from 12.4 ms to 3.8 ms. Hands assembled earlier keep the factory 500 µs until you run configure_motor_chain --return-delay. - A motor that is rebooted now gets its current limit, gains and profile written back once it answers again. Before, a recovered motor could run without its current cap. The move to the neutral pose also stays in the configured control mode, so the hand no longer goes limp on the way in and out. - The servos' own position gains and their speed and acceleration profiles can now be read and written, on Feetech hands as well since 0.5.1. Feetech speed limits are taken from each motor's model, because the wrist and finger motors differ by more than a factor of two: Changes.
Asimov v1
!Asimov v1
- The joint ranges in the URDF and the MuJoCo model now match the hardware limits. Hip pitch is 1.57 rad in each direction on both legs, where the model allowed 2.09 rad one way and 1.0 rad the other, and the elbows stop at 2.09 rad, down from 2.44. - A policy trained on the old model may ask for hip or elbow angles the real robot cannot reach, so retrain or check it against the new limits before running it on hardware: Changes.
Why it matters
Much of this week is about names and numbers that other code depends on. PX4 renamed its position topic, the HX5-D20 hand renumbered its joints, Upkie dropped its wrappers and Isaac Lab removed its install script, so pulling the latest version of any of them can break code that was working. The same goes for the models: Asimov v1 narrowed its joint ranges to what the hardware can do, and ROBOTIS and Wuji published simulation models from the people who build the robots. A policy is only as good as the description it was trained on, so before you upgrade, compare the joint names, limits and topics your code uses with the new ones, and plan to retrain where a limit has moved.