Open Source Robotics

A curated collection of high-quality open source robotics projects, tools, and software to propel the robotics community forward.

Project highlights: 9 to 15 November

  • Source: Open Source Robotics
  • Published: 2026-11-16
  • Category: upstream

mjlab fresh installs broke MuJoCo, MuJoCo 4.0.1 fixes a 4.0.0 solver bug, Rerun 0.43 and Strands Robots 0.6.0 break existing code, and uni-robot-agent changes its protocols.

mjlab

!mjlab

- Version 1.6.1 fixes fresh installs that came out with a broken MuJoCo. A new release of the viewer package mjviser (0.0.17) depends on mujoco-core instead of mujoco, so pip and uv installed mujoco-core 4.0.1 beside the pinned mujoco 3.11, and both wrote into the same mujoco/ folder. - Older mjlab releases are fixed by adding "mjviser<0.0.17" to your dependencies; 1.6.1 caps it the same way. Release notes - Why it matters: if you installed or rebuilt an mjlab environment this week and MuJoCo misbehaves or fails to import, the cause is this clash, not your code; add the cap, or move to 1.6.1, and reinstall into a clean environment.

MuJoCo

!MuJoCo

- Version 4.0.1 fixes a bug introduced in 4.0.0 in the sparse Cholesky factorisation. With a sparse constraint Jacobian, the Newton solver was given a wrong Hessian factor, which showed up as slow convergence, wrong accelerations and divergence. - In the 4.0.0 macOS and Linux archives, one dynamic library that failed to load (the USD decoder plugin) aborted simulate at startup; it is now skipped with a warning. Release notes - Why it matters: if you ran models on MuJoCo 4.0.0 and saw a solver diverge or accelerations that looked wrong, update to 4.0.1 and re-run before you trust those results or train on them.

Rerun

!Rerun

- Code that embeds the web viewer from JavaScript, or sets up storage on Rerun Hub, breaks on 0.43: the JavaScript API now uses camelCase and object_store_config is renamed storage_location (StorageLocationAlias in Python and Rust). Version 0.43.1 accepts snake_case web viewer options again. - New in 0.43: you can edit the translation of grid maps and volumes, and the voxel size, in the viewer; memory limits now respect Linux cgroup limits; and 0.43.2 fixes points and depth clouds not drawing in Safari. Release notes - Why it matters: if you embed the Rerun web viewer from JavaScript or configure storage on Rerun Hub, use 0.43.1 or later and follow the migration guide; if you run Rerun in a container, its RAM limit now follows the container's limit instead of the whole machine's.

Strands Robots

!Strands Robots

- Version 0.6.0 is a rewrite around three classes: Simulation (a world, on MuJoCo or on GPU engines), Robot and Policy. The same code now drives a simulated robot, a real one or one on another machine, and 76 robot presets load by name. - The 0.5.x API is removed. Out-of-range commands from an agent are now refused with an error rather than clipped. Release notes - Why it matters: if you wrote code against Strands Robots 0.5.x, it will not run on 0.6.0 until you port it, so pin 0.5.x until then; if an agent drives your robot, expect refusals where it used to be silently clamped.

uni-robot-agent

!uni-robot-agent

- Version 0.2.0 adds a "clean mode" of four tools (move, hold, capture_scene, finish), an opt-in window that drops old inline images from the observation history, an offline doctor command and a real-robot integration for the ARX ONE. - It moves to environment protocol 4, tool schema 8 and actor HTTP protocol v3; older services are not compatible. Native environments need uni-robot-agent-runtime 0.2.0 installed separately, and the two packages must never share a Python environment. Release notes - Why it matters: if you run uni-robot-agent clients and environment services on different machines, upgrade them together or they will stop talking; if you install both packages into one environment, you will break it.

AI Worker

!AI Worker

- Release 2.3.3 detects a disconnected leader arm through configurable heartbeat timeouts, hands control back to external control when default-style teleoperation stops, waits for the initial poses to finish before teleoperation starts, and publishes gripper actions in both joint and end-effector modes. Release notes - After 2.3.3, the F1 and F2 robots were renamed W1 and W2 on the main branch: launch files, config folders and the MoveIt model:= argument (w1, w2) all change. Changes - Why it matters: if you teleoperate an AI Worker, a disconnected leader arm is now detected, with timeouts you can set; and any script or launch command of yours that names f1 or f2 will fail once you move to a version with the rename.

Microduck

!Microduck

- Version 0.21.0 lets you switch the robot's mode by touching an NFC tag to its head: walk, roller skating, laser following, soccer, and an autonomous mode in which the duck wanders, naps, chases a laser dot, plays with a ball and maps the room. robotctl nfc-flash writes the tags. - Version 0.21.1 adds a way into the robot over the network: hold the face button for 5 seconds and SSH opens on wifi for 30 minutes. The protocol goes from API 55 to 61, all additive. Release notes - Why it matters: update all the daemons on the robot together, because a v59 robotd refuses the new hello {client} and robot.head {alpha} calls by name; and note that the maintainers released 0.21.0 and 0.21.1 directly, checked in CI but not yet run on a robot, so try them on the bench first.

ORCA Hand

!ORCA Hand

- Orca Core 0.13.0 adds guided tensioning (scripts/guided_tension.py) for v2.1 hands: the motors hold lightly while you move each joint through its range, the script scores how taut each tendon is, then holds them while you ratchet the ones that fall short. v2.0 hands still use scripts/tension.py. - It also adds manual calibration where you move each joint to its stops by hand, hardware versions reported as three digits (210 for v2.1.0), and a touch-sensor fix: a busy program could previously lose touch until it reconnected. Changes - Why it matters: if you build a v2.1 hand, tension it with the guided script and calibrate last, and avoid hard pushes on a finger while the motors hold, as that can loosen the tendon; calibration records now save beside the config you load, so with more than one hand of the same model give each its own copy of the config.

SO-ARM 102

!SO-ARM 102

- The assembly guide now follows booklet v0.8 and names the servo for each joint: all leader servos are STS3215; on the follower, IDs 1 and 4 are Feetech STS3235 and IDs 2 and 3 are STS3250. - Screws marked "servo kit" on a step come from the servo pack, not the loose fastener bag. The board holder and cap are renumbered SO102.02.101, and the camera holder step now takes five M3×10 screws instead of four. Changes - Why it matters: if you are buying servos or printing parts for an SO-ARM 102 follower, check the servo model for each ID against the v0.8 guide before you order, and print the v0.8 booklet rather than v0.7, because part numbers and screw counts differ.

pib

!pib

- Version 0.7.12 puts imitate_hands back in the right direction: the sign of the right lower-arm rotation is inverted again and the hand-tracking crop is fixed. - Version 0.7.14 changes how pib installs and updates: setup-pib.sh now checks out the newest release tag rather than a branch, updates pull prebuilt images from the registry (with a local build as fallback), and free disk space is checked before the update touches git. Version 0.7.17 turns camera stereo depth off by default. Release notes - Why it matters: a pib that copies a person's arm movements was moving the right arm the wrong way until 0.7.12, so update before you demonstrate it. After moving to 0.7.17, depth-based programs should be tested, since depth is no longer on unless you enable it.

Why it matters

Four of this week's changes can stop working code. mjlab users who installed or rebuilt an environment this week may have a broken MuJoCo and need the mjviser cap. Rerun 0.43 renames JavaScript options and a storage setting, and Strands Robots 0.6.0 removes the 0.5.x API, so both need porting before you update. uni-robot-agent 0.2.0 needs clients and services upgraded together. MuJoCo 4.0.0 users should update to 4.0.1 and re-run models that diverged. Microduck's 0.21 releases have not yet been run on a robot, so try them on the bench first.

Related projects

  • mjlab
  • MuJoCo
  • Rerun
  • Strands Robots
  • uni-robot-agent
  • AI Worker
  • Microduck
  • ORCA Hand
  • SO-ARM 102
  • pib

This website requires JavaScript to function properly. Please enable JavaScript to explore our collection of open source robotics projects.