View a markdown version of this page

Train a robot manipulation policy with MuJoCo on Deadline Cloud - Deadline Cloud

Train a robot manipulation policy with MuJoCo on Deadline Cloud

The mujoco_sim_to_policy job bundle trains and renders a learned robot-manipulation policy on Deadline Cloud using a MuJoCo simulation of the Strands Robots so100 arm. The job runs three dependent steps on managed GPU workers:

  1. Datagen — Scripted joint-space pick of a cube in MuJoCo, recorded as a LeRobot dataset. Each episode is verified by a cube-height check and discarded if it fails.

  2. Train — Finetunes a LeRobot ACT policy on the generated dataset (CUDA).

  3. Render — Drives a MuJoCo rollout with the finetuned policy and records the result as an MP4 video and PNG frames.

The bundle requires a Linux x86_64 GPU fleet (steps render headless through EGL and train on CUDA) and a conda queue environment that consumes the CondaPackages and CondaChannels job parameters. The default conda packages are python=3.12 pip git ffmpeg on conda-forge.

From the job_bundles directory, submit the job:

OUT="$(pwd)/output"; mkdir -p "$OUT"
deadline bundle submit mujoco_sim_to_policy -p "OutputDir=$OUT" --yes