harryzhang1018/TestRandomization

Test randomization for ground vehicle controllers: data

This dataset accompanies the paper Screening Ground Vehicle Controllers via Test Randomization: Sim2Real and Sim2Sim Studies. It contains:

1. code/: the scripts that produce the results of the paper, with intermediate files, the Case 2 policies, and the result records. How to reproduce each table: code/README.md.
2. videos/: videos of physical runs of the 1/6-scale vehicle on Paths 2 and 3.
3. simulation_logs/: raw simulation logs of the scale-vehicle study (Case 1) in the Chrono digital twin.

The logs of all 60 physical runs and the Docker image with the controller source are at https://github.com/uwsbel/sbel-reproducibility/tree/master/2024/PathFollowingSim2real.

Videos

One physical run of each controller on Paths 2 and 3. There is no video of the Path 1 runs.

File	Path	Controller	Duration (s)
videos/path2_mpc.mp4	2	MPC	55
videos/path2_nnmpc.mp4	2	NN-MPC	55
videos/path2_pid.mp4	2	PID	56
videos/path2_nnhd.mp4	2	NN-HD	56
videos/path3_mpc.mp4	3	MPC	176
videos/path3_nnmpc.mp4	3	NN-MPC	154
videos/path3_pid.mp4	3	PID	157
videos/path3_nnhd.mp4	3	NN-HD	173

The physical logs of Path 2 are in TestingData/syn_path and those of Path 3 in TestingData/lot17 of the GitHub repository.

Simulation logs

One archive per batch (35,144 episodes in total). Extract with tar -xzf <archive>.

Archive	Content	Episodes
case1_B01_main_tr.tar.gz	Recovery tests, offset range S2 (400 starts), 40 s horizon; metrics also for 15 s and 25 s	2,400
case1_B02_c0_arm.tar.gz	As B01, with control starting 0.5 s after the episode start	2,000
case1_B03_nominal.tar.gz	Nominal simulation on Paths 1-3, five controller phase offsets per path	90
case1_B04_rf_a.tar.gz	Monte Carlo on Paths 1-3: start jitter of up to 0.2 m in x and y and 0.067 rad in yaw, random controller phase; 20 runs per path	300
case1_B04w_rf_a_wide.tar.gz	As B04, with a wider start jitter	300
case1_B05_rf_b.tar.gz	As B04, with condition D5 (delay, steering bias, pose noise)	300
case1_B06_severity_shape.tar.gz	Recovery tests for offset ranges S0, S1, S3, S4, lateral-only and heading-only ranges, and a Beta(3,3) shape; 200 starts each	9,000
case1_B07_dist_recovery.tar.gz	Recovery tests (S2, 200 starts) under conditions D1a-D5	9,000
case1_B08_dist_fullpath.tar.gz	Monte Carlo on Paths 1-3 (10 runs per path) under conditions D1a-D4b	1,200
case1_B09_placement_floor.tar.gz	Recovery tests (S2, 50 starts) at five placements of the reference line in the world frame	1,250
case1_B10_sensitivity.tar.gz	Recovery tests (S2, 200 starts) with ideal latency, with a 0.095 s controller period, and with default OSQP settings (MPC, two repeats)	2,400
case1_B11_ws7_tuning.tar.gz	Gain selection for Stanley and pure-pursuit controllers (S2 starts and two circular references)	4,356
case1_B12_ws7_eval.tar.gz	Stanley, pure pursuit, MPC-N5, MPC-sluggish, and PID-aggressive: recovery tests (S2, 400 starts), nominal simulation, and Monte Carlo under D5	2,225
case1_B13_determinism_audit.tar.gz	Repeats of 323 episodes of batches B01-B11; each repeat is byte-identical to its source episode	323

Controllers: MPC, NN-MPC, PID, and NN-HD with steering gains x1.0, x1.5, and x1.75; batches B11-B13 add the controllers listed above. The field tags.ctrl_desc of each episode names its controller.

Offset ranges of the recovery tests: S2 is |e| ~ U[1.5, 2.5] m and phi ~ U[-pi/4, pi/4]; the other ranges are given by tags.family and the initial pose in each episode configuration.

Disturbance conditions (tags.condition):

Condition	Actuation delay (s)	Steering bias	Pose noise SD (m, rad)	Friction
none	0	0	0, 0	0.9
D1a / D1b	0.10 / 0.20	0	0, 0	0.9
D2a / D2b	0	0.013 / 0.026	0, 0	0.9
D3a / D3b	0	0	0.02, 0.01 / 0.05, 0.03	0.9
D4a / D4b	0	0	0, 0	0.6 / 0.4
D5	0.10	0.013 (either sign)	0.02, 0.01	0.9

Steering bias is in normalized steering units.

Episode files

Each episode is a folder <batch>/episodes/<episode_id>/:

File	Content
episode_config.json	Initial pose, reference, controller, latency model, perturbation, horizon, and tags (batch, offset range, condition, path, controller description)
summary.json	Outcome, time of first entry into the reference tube, per-run mean absolute lateral and heading errors, metrics for each horizon (by_H), divergence flags, config_hash, and SHA-256 of the data files
state100.csv.gz	Vehicle state at 100 Hz: position, yaw, speed, applied inputs, and lateral and heading errors
ticks.csv.gz	Controller ticks: observation, error state, and raw and published commands
mpc_diag.csv.gz	MPC solver diagnostics (MPC episodes only)
obs_noise.csv.gz	Pose-noise draws (episodes with pose noise only)
RigidPlane_mu_*.json	Terrain file (episodes with reduced friction only)

Each batch also has summary.csv and summary.jsonl, with one row per episode.

config_hash is the SHA-256 of episode_config.json written as compact JSON with sorted keys. SHA-256 sums of the archives are in simulation_logs/SHA256SUMS.txt.

License

Data (videos, simulation logs, result records): CC BY 4.0. Code in code/ and the four checkpoints in code/case2/policies/released/: BSD 3-Clause, as in the original repository.