Which Randomizations Matter? A Seeded Ablation for Blind Hexapod Locomotion in MuJoCo Warp

Preprint reporting a seeded ablation of domain randomization for a blind (proprioceptive) 18-joint hexapod trained with PPO in MuJoCo Warp on a procedurally generated heightfield. Each policy trains in about 20 minutes on a laptop GPU. We ablate five training components (terrain diversity, friction, pushes, one-step action delay, observation noise) with three seeds each, and evaluate all 18 policies under paired conditions in 24 settings, including stress tests beyond the training distribution. Main findings: action-delay randomization has the clearest effect (removing it lowers linear tracking by 0.18 on average and by 0.36 when every episode is delayed). Observation-noise randomization improves yaw tracking only. Terrain diversity helps mainly on slopes beyond the training range. Friction and push randomization show no measurable benefit in the ranges tested, which with three seeds excludes only benefits larger than about 0.05. The study is simulation-only, with three seeds and one robot. Code: https://github.com/UpstairsEngineer48/Heaxpod_RLData (checkpoints, TensorBoard logs and per-run configs for all 18 runs): 10.5281/zenodo.24356733

Authors

Institutions

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-19
DOI
https://doi.org/10.5281/zenodo.24357266
Primary Topic
Reinforcement Learning in Robotics
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Which Randomizations Matter? A Seeded Ablation for Blind Hexapod Locomotion in MuJoCo Warp

Pranjal Pandey
Zenodo (CERN European Organization for Nuclear Research)
Reinforcement Learning in Robotics
preprint

Which Randomizations Matter? A Seeded Ablation for Blind Hexapod Locomotion in MuJoCo Warp

Pranjal Pandey
preprint en

Abstract

Preprint reporting a seeded ablation of domain randomization for a blind (proprioceptive) 18-joint hexapod trained with PPO in MuJoCo Warp on a procedurally generated heightfield. Each policy trains in about 20 minutes on a laptop GPU. We ablate five training components (terrain diversity, friction, pushes, one-step action delay, observation noise) with three seeds each, and evaluate all 18 policies under paired conditions in 24 settings, including stress tests beyond the training distribution. Main findings: action-delay randomization has the clearest effect (removing it lowers linear tracking by 0.18 on average and by 0.36 when every episode is delayed). Observation-noise randomization improves yaw tracking only. Terrain diversity helps mainly on slopes beyond the training range. Friction and push randomization show no measurable benefit in the ranges tested, which with three seeds excludes only benefits larger than about 0.05. The study is simulation-only, with three seeds and one robot. Code: https://github.com/UpstairsEngineer48/Heaxpod_RLData (checkpoints, TensorBoard logs and per-run configs for all 18 runs): 10.5281/zenodo.24356733

Zenodo (CERN European Organization for Nuclear Research)
National Institute of Technology Rourkela (IN)
Reinforcement Learning in Robotics
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon