> ## Documentation Index
> Fetch the complete documentation index at: https://docs.godrift.ai/llms.txt
> Use this file to discover all available pages before exploring further.
# Dual-arm — Trossen Aloha
> Set up an Aloha bimanual rig on a tabletop and reach both arms toward objects in one prompt
The Trossen Aloha is the most-cited open bimanual platform in 2026 manipulation research — easy to recognize, well-documented, and present in [MuJoCo Menagerie](https://github.com/google-deepmind/mujoco_menagerie) with correct joint limits, gripper kinematics, and inertias.
Drift composes the scene, the objects, the dual-arm controller, and the sensor logging around the Menagerie Aloha model.
Recommended setup: MuJoCo 3.8.1+, Python 3.10+. Drift can clone Menagerie for you on the first run if it isn't on disk yet.
## The prompt
```drift theme={null}
set up a Trossen Aloha bimanual scene from mujoco_menagerie on a
tabletop (1.0×0.6m at z=0.75m), place three 0.05m cubes at
table-local (0.2, 0.15, 0.025), (0.2, 0, 0.025), (0.2, -0.15, 0.025)
in red/green/blue, write a Python control loop that moves the left
end-effector to hover 5cm above the leftmost cube and the right
end-effector to hover 5cm above the rightmost cube using
joint-position interpolation from the default keyframe, run 5s
headless, log both wrists' force/torque (Drift adds the sensor sites
to the Aloha if they aren't there already) and all 14 joint angles
```
From this single prompt, Drift:
* Locates or clones `mujoco_menagerie`
* Composes `aloha_scene.xml` that ``s the Menagerie Aloha and adds the table + three cubes at the requested colors and positions
* Writes joint-space targets per arm that approximate the wrist hovering over each target cube. The first attempt comes from MuJoCo's forward kinematics — pose, check wrist position, nudge — but it's not a full IK solver and may overshoot or undershoot by a few centimetres. Refine with a follow-up like `"the left wrist is 5cm short — adjust the shoulder-pitch target by +0.1 rad"`.
* Writes a `run_sim.py` that linearly interpolates between current and target joint angles over 5 seconds, applies position control, and logs at 100 Hz
* Runs headless, prints the final EE positions vs targets, and notes any contact between wrist and table
## What ends up on disk
```
your-project/
├── aloha_scene.xml # table + 3 cubes + Menagerie Aloha include
├── run_sim.py # bimanual joint-space interpolation
├── sensor_log.csv # 500 rows: 14 joints, 2× F/T sensors
└── DRIFT.md # scene summary, target poses, what to iterate on
```
## What the viewer shows
```drift theme={null}
open aloha_scene.xml in the MuJoCo viewer
```
Both Aloha arms move smoothly from their start pose. The left wrist tracks over the red cube; the right wrist tracks over the blue cube. The middle (green) cube stays untouched. The F/T traces in `sensor_log.csv` are near-zero through the motion (free space) and would spike if the arms touched the table.
## Iterating from here
Where the demo gets interesting is the follow-ups:
```drift theme={null}
close both grippers once each EE is over its target cube and lift
2cm — log gripper force during contact
add three larger 0.08m cylinders behind the cubes — make the right
arm reach the rear-right cylinder while the left does the red cube
add a top-down RGB camera over the center of the table and write
30fps frames to a video file
load an inverse kinematics solver and switch the controller to
Cartesian EE targets instead of joint targets
swap to Mujoco-MPC and add an end-effector position cost on each arm
add an Aloha-shaped ROS2 bridge: /aloha/left/joint_states,
/aloha/right/joint_states, /aloha/left/cmd, /aloha/right/cmd
```
## Honest scope
Things this guide does:
* Compose the bimanual scene around the Menagerie Aloha model
* Write a per-arm joint-space interpolation controller
* Wire both wrist F/T sensors and the 14 joint angles, log them
* Open the viewer / run headless
Things this guide does **not** do:
* A learned bimanual grasp policy (ALOHA Unleashed–style behaviour cloning) — those are policies you train; Drift can scaffold the data-collection environment but not the policy
* Full collision-aware motion planning between the arms
* Tactile-sensor-driven manipulation — Aloha's grippers don't carry tactile sensors in the Menagerie model
If you have a policy from a training run, ask Drift to wire it in:
```drift theme={null}
load /path/to/aloha_cube_grasp_policy.onnx and run it at 50Hz,
feeding joint positions + RGB camera images as observation,
outputting target joint positions for both arms
```
## Common gotchas
The table needs `condim=3` (or higher) on its contact and a non-zero friction. The default plane material in some Menagerie scenes is frictionless. Tell Drift `"the cubes fall through the table — make the table-cube contact realistic"` and it'll fix the `` block.
Position-control PD gains are model-specific. Aloha's default `kp` in Menagerie is conservative. Ask Drift to ramp the gains:
```drift theme={null}
bump the position-control kp on every Aloha joint to 1.5× and re-run,
also damp by 20%
```
Aloha's right arm has a different joint convention than the left in the original ROS package — but Menagerie normalises this. If you see overshoot only on one side, double-check the sign convention in your interpolation. Drift will catch this if you ask `"why is the right arm overshooting"`.
## Next steps
Locomotion-shaped showcase
Full-body PD-balanced demo
The base MuJoCo guide — useful background for everything in Showcase
Capture your manipulation conventions (IK solver pick, gripper-force rules) as a skill