Interpreting Residual Control Gains in Quadrotor Simulation - simulation records and code

Description

Compact numerical records, fixed configurations and source code for the simulation study "Interpreting Residual Control Gains in Quadrotor Simulation". The study distinguishes available residual corrections, fixed-budget SAC learning, classical-controller comparisons, force-application checks and failure-cost accounting. All adverse outcomes are retained. S1 contains the historical two-implementation comparison and verification code. S2 and S3 provide descriptive displays of those records. S4 contains the distinct 1,920-episode fixed-policy parameter holdout. S5 contains holdout trace accounting and a 480-episode targeted force-path intervention. S6 contains the complete corrected-platform comparison: six classical families, five SAC training seeds, 400 checkpoint-development evaluations and 17,280 confirmation evaluations, including all 5,038 failures and independent reaggregation code. The archives preserve their original bytes and identifiers. Historical PyBullet results use the legacy force path; S6 uses the corrected path. The independent rigid-body implementation uses RK4 integration. These are simulation results for one nominal vehicle and finite development budgets, not real-flight validation. Full high-frequency trajectories, learning checkpoints and replay buffers are not included in this compact deposit; they remain in the authors archive. The deposited compact records support the reported aggregate comparisons. See README and the SHA-256 manifest for file scope and verification instructions.

Steps to reproduce

1. Verify the SHA-256 hashes in MANIFEST.json against the six original ZIP files. 2. Unpack each archive into a separate directory and follow the archive-specific README and verifier. Preserve the historical archive identifiers. 3. For the corrected-platform study, unpack Supplementary_Data_S6.zip and run: python verify_compact.py --output-dir /your/fresh/verification-directory 4. The compact verifier uses Python, NumPy and SciPy; tested versions are Python 3.12.12, NumPy 2.4.1 and SciPy 1.17.0. It checks hashes, selections, 72 method means, 864 condition cells, eight sign tests and five-seed summaries, without simulation or training. 5. Full high-frequency trajectories and training checkpoint/replay arrays are outside this compact deposit. Raw-control audit receipts document their earlier checks; compact reproduction cannot repeat those checks from absent arrays. See the README for these boundaries and third-party licence notices.

Institutions

• Nanchang Hangkong University

  Jiangxi, Nanchang

Categories

Licence