Code for Behavioral Denoising: A New Perspective on Nonlinear Behavioral Control Inspired by Closed-Form Diffusion Policies.
This branch contains the current implementation and runnable examples for
Cartpole, Rocket Landing, Reacher, and Push-T.
For the L4DC submission code, saved results, and paper reproduction instructions,
see the l4dc branch.
Configurations · Implementation · Demonstrations · Video gallery
Selected rollouts from the L4DC submission, at real-time playback. Each panel holds its final pose after the run ends. The updated Push-T controller is shown in the gallery below.
Comparison with Select-DPC
BD-χ² (top, green) and Select-DPC (bottom, orange), starting from the same states. These selected submission trials show BD-χ² succeeding where Select-DPC fails. Playback is synchronized by simulation time, with final poses held after termination.
BD-χ² and Select-DPC now express the pusher position and planned action targets in the block's coordinate frame, while retaining the block's absolute pose. All rows below use the same 50 starts, demonstrations, and execution schedule. World-coordinate rows are the previous results. Controller parameters were not retuned for the relative representation.
| Method | Coordinates | Mean peak coverage | Mean final coverage |
|---|---|---|---|
| BD-χ² | World | 63.09% | 57.99% |
| BD-χ² | Block-relative | 94.43% | 88.16% |
| Select-DPC | World | 6.28% | 2.23% |
| Select-DPC | Block-relative | 10.55% | 0.00% |
Relative Select-DPC produced frequent OSQP inaccurate-solution warnings. These values include all trials and the controller's existing fallback behavior.
Watch all 50 updated trials, with peak and final coverage for each trial. Coverage is the fraction of the goal shape covered by the block. Trials stop above 95% coverage or after 30 seconds. The L4DC submission gallery retains all 50 original rollouts.
Tested with Python 3.10 on Ubuntu 22.04. From the repository root, create separate environments for the control benchmarks and Push-T, which require different NumPy versions:
python3.10 -m venv .venv
.venv/bin/python -m pip install -r requirements/benchmark.txt
python3.10 -m venv .venv-pusht
.venv-pusht/bin/python -m pip install -r requirements/pusht.txtThe simulators and demonstrations are included. No GPU is needed.
Use one numerical thread per worker. If a robotics shell such as ROS sets
PYTHONPATH, unset it first: imports use src.*, so src/ itself must not be on
PYTHONPATH.
unset PYTHONPATH
export OPENBLAS_NUM_THREADS=1 OMP_NUM_THREADS=1 MKL_NUM_THREADS=1
export SDL_VIDEODRIVER=dummy MPLBACKEND=AggStart with one method and one initial state:
.venv/bin/python -m experiments.cartpole --methods bd_chi2 --seeds 10003 --output runs/check
.venv-pusht/bin/python -m experiments.pusht --methods bd_chi2 --seeds 300031 --output runs/checkTo run the configured methods and starts for each task:
.venv/bin/python -m experiments.cartpole --output runs/benchmark
.venv/bin/python -m experiments.rocket --output runs/benchmark
.venv/bin/python -m experiments.reacher --output runs/benchmark
.venv-pusht/bin/python -m experiments.pusht --output runs/benchmarkRecords are saved under <output>/<task>/<method>/seed_<seed>/, including states,
actions, references, outcomes, and timing. Repeat a command to resume it. Use a new
output directory after changing settings. Generated runs are not committed.
--methods accepts bd_mean, bd_sample, bd_kl, bd_mean_chi2, bd_chi2,
select_dpc, windowed_deepc, and deepc. Push-T supports bd_chi2 and select_dpc.
Use --steps for a shorter trial. DeePC can take many hours to run.
Both Push-T methods default to block_relative observations and action targets.
To select this representation explicitly, run:
.venv-pusht/bin/python -m experiments.pusht --representation block_relative \
--methods bd_chi2 select_dpc --output runs/pusht_relativeUse --representation world for both methods to use absolute coordinates.
The gallery above shows BD-χ². See the
coordinate configuration for per-method
overrides. Saved trajectories and executed actions remain in world coordinates.
Reports and visualizations read saved trajectories without rerunning controllers:
.venv/bin/python scripts/report.py --root runs/benchmark
.venv/bin/python scripts/plot.py --root runs/benchmark --group control
.venv-pusht/bin/python scripts/plot.py --root runs/benchmark --group pushtAdd --videos and install FFmpeg to generate motion videos. See the
gallery instructions to build a Push-T page from your own runs
or preview the published gallery assets. Paper figures and their reconstruction
scripts are available on l4dc.
Timing replays saved controller queries and excludes initialization and simulation:
.venv/bin/python scripts/timing.py --root runs/benchmark --task cartpole --method bd_chi2The timing record includes process CPU time, CPU affinity, SMT topology, and background load. Use a quiet machine for comparisons.
Run tests in their corresponding environments. No saved benchmark results are required:
.venv/bin/python -m pytest -q -rs --ignore=tests/test_pusht.py --ignore=tests/test_pusht_task.py --ignore=tests/test_select_dpc_chunks.py
.venv-pusht/bin/python -m pytest -q -rs tests/test_pusht.py tests/test_pusht_task.py tests/test_pusht_coordinates.py tests/test_pusht_representations.py tests/test_select_dpc_chunks.py tests/test_gallery.pyWe thank the authors of the following awesome projects:
- Select-DPC, from which we adapted the control benchmarks and the Select-DPC, DeePC, and Windowed DeePC baseline implementations.
- Diffusion Policy, for the Push-T environment and released human demonstrations.
- COCO Rocket Lander, for the Rocket simulator and dynamics model.
Simulation uses MuJoCo, Gymnasium, Box2D, and Pymunk. The MPC experts and optimization baselines use CasADi, IPOPT, CVXPY, and OSQP.
We used Codex for assistance with code development, review, testing, and documentation.
The project is released under the MIT license. See CITATION.cff and third-party notices.


