Hunter Locomotion on Agricultural Soil
PPO policies for the Hunter biped, trained in IsaacSim / IsaacLab with HumanoidVerse, for walking on deformable, furrowed agricultural terrain (DFH soil model). Hunter-only.
What to use
walkers/mildsoil_walker/model_5050.pt — recommended deployment policy: a
forward walker on mild soil (~0.12 BW drag).
| Folder | Description |
|---|---|
walkers/mildsoil_walker |
Forward walker on soil — deploy this |
walkers/v7_baseline |
Flat-ground walker (warm-start root; shown above) |
walkers/rigid_walker |
Walker on rigid furrows |
dfh_chain/ |
12-stage drag-survival curriculum — research / warm-start only |
Usage
Each folder ships its training config.yaml, so a checkpoint loads directly:
python humanoidverse/eval_agent.py +checkpoint=walkers/mildsoil_walker/model_5050.pt
Algo: PPO-ROA · Obs: leggedloco_obs_history_wolinvel (234-dim) ·
Sim: IsaacSim 4.2 · License: MIT
Fuzzy-curriculum ablation (Set B)
| fuzzy_curriculum/ (9 policies) | Set B adaptive-vs-fixed friction-curriculum ablation (rigid PhysX): fixed/crisp/fuzzy x 3 seeds. Fuzzy curriculum ~= crisp (no significant gain). | Research |
Fuzzy soil-index finding: across rigid (Set A/B) and DFH deformable (Set C) soil, the paper's fuzzy Mamdani index adds no demonstrable value over raw traction/support — even where the support axis is physically active (Set C: support CV-R^2 +0.33 vs fuzzy +0.025). The failure is structural, not the simulator. See the code repo's
docs/experiments/fuzzy_soil_UNIFIED_conclusion.md.