FROM CONTRACTION TO CHAOS: COMPOUNDING ERROR RESISTANT WORLD MODELS FOR ROBOTS

This anonymized repository accompanies our paper under review at ICLR 2027 on OpenReview: https://openreview.net/forum?id=IKsP4J04TN

Standalone, anonymized release of the source code, configurations and data required to reproduce the experiments of the paper (ICLR 2027 submission). It is a consolidated extract of a larger research codebase: only the code paths, hydra configurations and dataset files exercised by the released experiments are included, with their logic and naming kept identical.

Note: the paper initially targeted ICRA, so several directory / experiment names still carry the ICRA2026 / RLRP-757 tokens. Likewise a few shared helper packages keep their historical names from the parent codebase (e.g. tools/f110_gym_env_tools/plot_tensorboard.py, the tensorboard writer used by every pipeline); no simulator other than the Lorenz attractor and the three csv datasets is included.

Contents

MTM-Pro/
├── docker/                    Dockerfile, docker-compose.yaml, requirements.txt (the ONLY supported runtime)
├── scripts/                   run_experiment.bash (paper runs), smoke_test.bash (fast end-to-end check)
├── src/
│   ├── launcher/              hydra entry points + configs (see launcher/configs/multirun_overrides/)
│   ├── algorithm/             experience-replay learning loop, motion-model learning/deployment
│   ├── pipeline/              Lorenz (math_env) and robotic (robotic_3d_env) experiment pipelines
│   ├── tools/                 models (tools/multistep_tools/models: MTM-Pro + baselines), data, plotting
│   ├── simulator/, custom_types/
├── data/
│   ├── external_data/         the three robotic datasets (csv trajectories used by the paper configs)
│   │   ├── long_horizon/data/neurobem   UAV NeuroBEM
│   │   ├── long_horizon/data/pi_tcn     UAV PI-TCN
│   │   └── ugv/husky_dataset_src/...    UGV Husky (gravel and grass, online split)
│   └── data_consolidated/     the numeric series behind every paper figure (csv + provenance meta.txt)
└── artifact/                  experiment outputs (created at runtime, VCS ignored)

The Lorenz attractor data is generated on the fly (math-gymnasium). The three norlab / fork python packages the code depends on (mbrl-lib fork, math-gymnasium, trajectory-container-tools) are pulled from their git repositories at build time (see pyproject.toml).

Quick start

Requirements: git, docker (>= 24, with the compose plugin). NVIDIA GPU runs additionally need the NVIDIA container toolkit.

git clone <this repository> MTM-Pro          # the directory MUST be named MTM-Pro
cd MTM-Pro

# 1. Build the image (~10-20 min, downloads PyTorch)
docker compose -f docker/docker-compose.yaml build

# 2. Fast end-to-end check (CPU, ~3 min): one MTM-Pro run on Lorenz + one on the UAV dataset,
#    with tiny models and 2 epochs
docker compose -f docker/docker-compose.yaml run --rm mtm-pro bash /opt/MTM-Pro/scripts/smoke_test.bash quick

# 3. A paper experiment (launch training, 4 seeds at the time, of one model on one dataset), on GPU
docker compose -f docker/docker-compose.yaml run --rm mtm-pro-gpu bash /opt/MTM-Pro/scripts/run_experiment.bash lorenz mtm-pro-ms+cp

Use the mtm-pro service instead of mtm-pro-gpu to run on CPU. Both services bind mount the repository at /opt/MTM-Pro, so outputs land in ./artifact/ on the host. An interactive shell:

docker compose -f docker/docker-compose.yaml run --rm mtm-pro-gpu   # working dir: /opt/MTM-Pro/src

Running the paper experiments

Every experiment is a hydra multirun (trial_nb=range(1,5), i.e. 4 model seeds) defined by ONE configuration file. scripts/run_experiment.bash <dataset> <model> [hydra overrides...] maps the paper labels to those files:

<dataset>	Configuration directory (src/launcher/configs/multirun_overrides/)	Launcher
lorenz	math_env/RLRP-757-E4b-low-global-noise	launcher/math_env_main.py
husky	robotic_3d_env/RLRP-757-Husky-F2-online	launcher/robotic_3d_env_main_multirun.py
neurobem	robotic_3d_env/RLRP-757-Neurobem-F1	launcher/robotic_3d_env_main_multirun.py
pi_tcn	robotic_3d_env/RLRP-757-PI-TCN-B1	launcher/robotic_3d_env_main_multirun.py

<model>	Paper label	Configuration file (robotic datasets / Lorenz)
ar-mlp-det	(AR) Det MLP	multirun-ar-mlp-det / math_env_main_multirun_RLRP-757-ar-mlp-det
ar-mlp-prob	(AR) Prob MLP	multirun-ar-mlp-prob / ..._RLRP-757-ar-mlp-prob
gru	(AR) GRU	multirun-gru / ..._RLRP-757-gru
lstm	(AR) LSTM	multirun-lstm / ..._RLRP-757-lstm
tcn	(AR) TCN	multirun-tcn / ..._RLRP-757-tcn
tcn-bf16	(AR) TCN-bf16	multirun-tcn-bf16 / ..._RLRP-757-tcn-bf16
e2e-tcn	(MS) E2E-TCN	multirun-e2e-tcn-F<horizon> / ..._RLRP-757-e2e-tcn-F1000
m3	(MS) M3	multirun-m3-F<horizon> / ..._RLRP-757-m3-F1000
mtm-pro-ms+cp	MTM-Pro-MS+CP (ours)	multirun-DMtm-Pro-MS+CP (+no-enf-continuity on PI-TCN) / ..._RLRP-757-DMtmPro-MS+CP
mtm-pro-ms	MTM-Pro-MS (ours)	multirun-DMtm-Pro-MS (+no-enf-continuity on PI-TCN) / ..._RLRP-757-DMtmPro-MS

The full list is in src/launcher/configs/multirun_overrides/ICLR2027_paper_release_experiments_list.md. The equivalent raw command of e.g. scripts/run_experiment.bash husky tcn is (from src/):

python launcher/robotic_3d_env_main_multirun.py --multirun \
    --config-dir=launcher/configs/multirun_overrides/robotic_3d_env/RLRP-757-Husky-F2-online \
    --config-name=multirun-tcn

Any hydra override can be appended, e.g. trial_nb=1 for a single seed, or hydra.launcher.n_jobs=2 (also HPO_N_JOBS=2 in the environment) to train two seeds concurrently on one GPU.

Outputs (tensorboard logs, checkpoints, test-time rollout records and plots) are written under artifact/ICRA2026/<experiment>/<date>/<time>/. The robotic pipelines cache the ingested dataset next to the csv files (<data_path>/ms_hdf5/ and <data_path>/ms_replaybuffer/) on first use.

Figures (test-time rollout comparison across models)

Once the multiruns of one dataset are recorded, the paper figures are rendered by the plot pipeline, which reads the recorded rollouts from artifact/ (one configuration per dataset in src/launcher/configs/multirun_testtime_rollout_plot/):

python launcher/multirun_testtime_rollout_plot.py multirun_testtime_rollout_plot@_global_=icra2026_Lorenz_RLRP-757-E4B-low-global-noise
python launcher/multirun_testtime_rollout_plot.py multirun_testtime_rollout_plot@_global_=icra2026_ugv_RLRP-757-F2c-obsD6-online
python launcher/multirun_testtime_rollout_plot.py multirun_testtime_rollout_plot@_global_=icra2026_neurobem_RLRP-757-F1
python launcher/multirun_testtime_rollout_plot.py multirun_testtime_rollout_plot@_global_=icra2026_pi_tcn_RLRP-757-B1a

Those configurations point to the experiment directories of the original runs (multirun_experiments_path): update them to the artifact/... paths of your runs. With consolidate_data.enable=true the pipeline also dumps the exact numeric series rendered in each figure to data/data_consolidated/<experiment>/ -- the version shipped here (data/data_consolidated/, see its README.md) is the one produced from the paper runs, so every figure can be rebuilt from csv without re-running the pipelines.

Smoke test / development configurations

scripts/smoke_test.bash [quick] runs the real experiment configurations of every model family on every dataset with the dev hydra overlay (src/launcher/configs/dev/): tiny models, 2 epochs, one seed and, for the robotic datasets, a handful of trajectories. It validates the environment, the data and every pipeline stage in minutes on CPU but is NOT a paper result. Any experiment can be run this way by appending, e.g., +dev@_global_=math_env_debug_mtm_pro ms_model.ensemble_size=1 trial_nb=1.

Datasets and credits

• UAV PI-TCN: the quadrotor dataset of Physics-Inspired Temporal Learning of Quadrotor Dynamics for Accurate Model Predictive Trajectory Tracking (arplaboratory/pi-tcn), pre-processed to csv. See data/external_data/Quadcopter_README.md.
• UAV NeuroBEM: the NeuroBEM quadrotor dataset (RPG, University of Zurich), pre-processed to csv. See data/external_data/Quadcopter_README.md.
• UGV Husky (gravel and grass, online split): our own Clearpath Husky A200 recordings, pre-processed to csv.
• Baselines: the baseline network implementations in src/tools/baseline_models/ derive from the long-horizon-dynamics codebase of Learning Long-Horizon Predictions for Quadrotor Dynamics (IROS 2024), see src/tools/baseline_models/README.md. Every model of the paper (MTM-Pro and the AR / MS baselines) is wrapped in src/tools/multistep_tools/models/.
• RedLeader962/mbrl-lib is a fork of Facebook Research's MBRL-Lib (MIT license).

License

See the LICENSE file of each third-party dependency (mbrl-lib fork, math-gymnasium, trajectory-container-tools) for the third-party code. The MTM-Pro source code is released for review purposes; a license will be attached to the de-anonymized release.