ksha23/nedm-study4-go2-crm

NeDM Study 4: a Go2 quadruped on CRM granular terrain

Everything needed to reproduce Study 4 of Learning the Right Abstraction: Neural Reduced Dynamics for Complex Robot Control (Zhang and Negrut, arXiv:2608.19375): the recorded corpora, the trained neural reduced models (surrogates), the fine-tuned policies, the raw Chrono scoring records, the base policy and the robot model, named by what they are.

A Unitree Go2 walks on Chrono's CRM terrain (SPH granular soil). A surrogate learned from recorded walking is used to fine-tune the robot's policy with PPO; every result is then verified in full Chrono, paired against the unmodified policy. With the recipe, forward tracking error on held-out paths falls 52% (ten of ten surrogates), yaw 67%, sideways 28%, with no regression on rigid ground.

• Code, recipe and step-by-step reproduction: https://github.com/uwsbel/NeDM, branch kyle/quadruped-pipeline, starting at quadruped/README.md. The job scripts that produced every result are in quadruped/jobs/.
• Write-up with every figure: https://claude.ai/artifact/G8PHCRfk7M8Mu7b8rW2PbU
• MANIFEST.tsv: every file here, its sha256 and size, the run name it had when it was produced, and the machine it came from.

Download

pip install -U huggingface_hub
hf download ksha23/nedm-study4-go2-crm --repo-type dataset --local-dir study4
# or one part, e.g. just the recipe's surrogates:
hf download ksha23/nedm-study4-go2-crm --repo-type dataset --local-dir study4 --include "surrogates/standard/*"

Using it with the code

From quadruped/ in the repository, with this dataset downloaded to study4/ and the corpora unpacked (commands below):

U=study4/assets/robot/go2_irrvis/urdf/go2_description.urdf
# fine-tune in one of the recipe's surrogates
python finetune.py --model study4/surrogates/standard/seed_8/best.pt \
  --policy study4/base_policy/policy.pt --corpus study4/corpora/normal_corpus/go2_crm_v2 \
  --out ft_s8 --method ppo --seed 0 --steps 100 --branches 1024 --iters 1000 --target-dw 1e9
# score any policy in Chrono on the 40 held-out paths, then pair against the base
python evaluate.py --policy ft_s8/policy_ft.pt --urdf $U --corpus study4/corpora/normal_corpus/go2_crm_v2 \
  --out ev/s8 --label s8 --terrain crm --paths 4 --seconds 15 --spawn-spread 1.0
# re-pair any published table from the raw records, no simulation needed
python paired_eval.py --base study4/results/hpcfund_c716f05e/eval_paths/base \
  --arms s8=study4/results/hpcfund_c716f05e/eval_paths/stop1k_h_s8_ms100_s0

Chrono itself (with FSI-SPH and GPU support) is built from the pinned source and patches described in the repository's quadruped/docs/STANDARD.md.

corpora/

Folder	What it is
normal_corpus/	go2_crm_v2: the recipe's corpus. 1,150 CRM episodes (920 train / 230 validation, split by episode), 2,041,752 rows at 100 Hz, about 5.7 h of walking. Collected on hpcfund (Chrono HIP build c716f05e). go2_crm_v2.tgz + manifest.json.
extended_corpus/	go2_crm_v23: the normal corpus plus 1,174 newly collected episodes, added as TRAINING data only, so the validation set is identical (2,094 train / 230 validation). Same build. go2_crm_v23.tar.zst + manifest.json.
rigid_corpus/	go2_rigid_v2: the same 24 shards x 50 episodes as the normal corpus (same seeds, commands, pushes and splits) on RIGID ground, for the soil-model control. Collected on sbel (build 3b0bd530). go2_rigid_v2.tar.zst + manifest.json.

Unpack: tar -xzf go2_crm_v2.tgz, zstd -dc go2_crm_v23.tar.zst | tar -xf -. Each corpus is a directory of per-segment CSVs plus a manifest recording the exact collection command, excitation and Chrono build.

base_policy/

policy.pt: the Go2 locomotion policy from rl_sar (policy/go2/robot_lab/policy.pt), unmodified; sha256 9f14cb95.... Every fine-tune starts from it and every result is paired against it.

surrogates/

Each <seed>/ holds best.pt (the selected checkpoint; train.py format, loadable by finetune.py and diagnostics/horizon_sweep.py) and metrics.jsonl (per-epoch training log). Unless stated, the architecture is 6 layers x 8 heads x 256 wide, context 128, and the state is the 36-D control-interface preset.

Folder	What it is
standard/	The recipe's surrogates, ten independent seeds: one-step training, then fine-tuned on their own 1 s rollouts. seed_6..9 trained on hpcfund, north_seed_0/1 on a lab desktop, euler_seed_2..5 on euler.
standard_one_step_only/	The same ten after stage 1 only (usable to about 2 s).
half_second_rollout/	The same ten fine-tuned on 0.5 s rollouts instead of 1 s (the horizon ablation; holds in 2 of 10). seed_7_rollout_count_study/ is the surrogate the rollout-count study used.
small/, large/	3x128 and 12x512 models on the normal corpus (model-size study).
standard_extended_corpus/	Standard models on the extended corpus.
large_extended_corpus/	12x512 on the extended corpus (no more accurate than standard; not fine-tuned).
quarter_corpus/, half_corpus/	Standard models on 25% and 50% of the normal corpus's training episodes.
channel_foot_forces/	40-D state: + the four foot normal contact forces.
channel_sinkage/	40-D state: + the four foot sinkages.
channel_forces_and_sinkage/	44-D state: both.
channel_3d_forces/	48-D state: + full 3-D foot contact forces.
rigid_ground/	Standard models trained on the rigid corpus (the soil-model control).
one_stage_16k/, one_stage_40k/	One-step and rollout losses together from scratch, 16k and 40k steps (no separate stage 1).

policies/

Each run folder holds policy_ft.pt (TorchScript, same interface as the base policy), finetune.json (settings and outcome, including the guard's verdict) and finetune.jsonl (per-iteration log), plus snapshots/ where the run saved them. Grouped by study:

Folder	What it is
recipe_iteration1000/	The headline result: the recipe at iteration 1000 in each of the ten standard surrogates (for seeds 6, 8, 9 and north 0 this is the iteration-1000 snapshot of the stopping study).
stopping_study/	Four runs to iteration 3000, snapshots every 500 iterations and at weight distances.
recipe_old_stop_dw4/	The recipe at the earlier weight-distance stop (dw 4, about 430 iterations).
rollout_count/	64 to 2048 parallel rollouts per update.
branch_length_and_ensembles/	0.30 s against 2 s branches; ensembles with and without a disagreement penalty.
disturbance_training/	robot_lab-style velocity kicks inside the surrogate (did not buy robustness).
ood_penalty_off/	The out-of-distribution penalty switched off.
induced_failures/	Raised learning rates, to calibrate the guard.
ppo_seed_test/	Other PPO seeds in seed 6's surrogate.
early_2s_ppo/	The first 2 s-branch PPO runs.
extended_corpus/	The recipe in the extended-corpus surrogates.
model_size/, less_data/	The recipe in the small, large, quarter- and half-corpus surrogates.
channels/	The recipe in the four channel-study surrogate families.
rigid_ground/	The recipe in the rigid-corpus surrogates.
one_stage/	The recipe in the one-stage surrogates.
analytic/	Analytic (backprop-through-the-model) fine-tunes by branch length (h15 = 0.30 s, h50 = 1 s, h100 = 2 s) and weight budget (dw), plus the PPO comparator.

results/

Raw paired-scoring records from evaluate.py, one folder per machine and Chrono build. Only records from the same folder may be paired (paired_eval.py refuses mixed builds). Each arm is a JSON list of episodes (per-axis tracking error, completion, uprightness) with a _manifest.json; base is the unmodified policy on that machine.

Folder	What it holds
hpcfund_c716f05e/eval_paths/	The 40 held-out CRM paths: the ten-surrogate table, channel study, capacity, one-stage, stopping study, and most ablations.
hpcfund_c716f05e/eval_h/	16 straight-walking CRM episodes, same arms.
hpcfund_c716f05e/eval_push_f*/	The push test at 120, 180, 240 and 300 N.
sbel_3b0bd530/	The soil-model control (eval_paths_sb on CRM, eval_rigid_sb on rigid ground) and the earlier rigid-ground checks (eval_rigid).
d33_53102025/	The analytic-vs-PPO grid (eval_paths_d33b; base in eval_paths_d33).
early_2026-09-21_22/	Results from the first days of the rebuilt pipeline (branch length, first 2 s runs), per machine.

To reproduce any table: python paired_eval.py --base <folder>/base --arms name=<folder>/<arm> ...

assets/

robot/go2_irrvis/: the Unitree Go2 robot description (URDF and meshes) every Chrono run loads; pass assets/robot/go2_irrvis/urdf/go2_description.urdf as --urdf.

Licenses

The corpora, surrogates, fine-tuned policies and scoring records are licensed CC-BY-4.0. Two third-party components keep their own licenses: the base policy (Apache-2.0, from rl_sar / robot_lab) and the robot model (BSD-3-Clause, from unitree_ros). See NOTICE.md.

Citation

@article{zhang2026nrd,
  title   = {Learning the Right Abstraction: Neural Reduced Dynamics for Complex Robot Control},
  author  = {Zhang, Harry and Negrut, Dan},
  journal = {arXiv preprint arXiv:2608.19375},
  year    = {2026}
}