Development - new features and changes over last months, mainly new TF training script - #66
Open
marcinpaluch1994 wants to merge 142 commits into
Open
Development - new features and changes over last months, mainly new TF training script#66marcinpaluch1994 wants to merge 142 commits into
marcinpaluch1994 wants to merge 142 commits into
Conversation
(cherry picked from commit 5f2cb2d)
…l on physical one.
…l on physical one.
…ller, along old trajectories.
…iment_3_04_2025 # Conflicts: # others/prepare_standard_experiment_folder.py
…neral state utilities.
![cursor[bot]](https://avatars.githubusercontent.com/in/1210556?s=60&v=4){kind=small}
There was a problem hiding this comment.
Cursor Bugbot has reviewed your changes and found 2 potential issues.
Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.
| high = np.array( | ||
| [ | ||
| self.theta_threshold_radians * 2, | ||
| self.angle_limit * 2, |
There was a problem hiding this comment.
Undefined angle_limit attribute crashes Cartpole_OpenAI initialization
High Severity
Cartpole_OpenAI.__init__ references self.angle_limit to build the observation space, but this attribute is never defined on the class. The old self.theta_threshold_radians was removed without replacing it with a self.angle_limit assignment. This causes an AttributeError crash whenever Cartpole_OpenAI is instantiated (e.g., when cartpole_type="openai" is selected in CartPoleEnv).
| s[POSITION_IDX+1]= vel | ||
|
|
||
| self.steps_beyond_terminated = None | ||
| return super().init_state(rng) |
There was a problem hiding this comment.
StabilizationOpenAI.init_state discards its constructed state
Medium Severity
StabilizationOpenAI.init_state carefully builds a near-upright state s with small angles and velocities (matching classic OpenAI CartPole specs), but then returns super().init_state(rng) which constructs and returns a different state with full-circle random angles. The locally built s is completely discarded, so episodes start with random angles instead of the intended near-upright configuration.
Point Control_Toolkit at the RPGD-C change that keeps OpenMP workers active between control ticks.
The top-level `import tensorflow` forced TF to load as soon as anything imported CartPole, i.e. before the controller/optimizer (and therefore the desired CPU/thread policy) was known. Move the import into create_rng so TF loads only when a TF RNG is actually requested, letting Driver/control.py set CPU affinity and TF/XLA threading before TensorFlow initializes its runtime. Co-authored-by: Cursor <cursoragent@cursor.com>
Watch config_controllers.yml and reload LQR gains live during a run; add a no-op controller_reset so safety-off works. Retune to Q=[10000,1,1,1], R=2000. Co-authored-by: Cursor <cursoragent@cursor.com>
Cast switch-timing config values to float and guard the random initialization so an infinite timeout yields inf instead of feeding inf into np.random.uniform. Co-authored-by: Cursor <cursoragent@cursor.com>
Update Control_Toolkit and SI_Toolkit submodule pointers to the verified working LQR state running on both PC and Zynq. Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Point the neural-imitator controller at CartPoleSimulation/SI_Toolkit_ASF/Experiments Dense-7IN-32H1-32H2-1OUT-8 (reliable swing-up + stabilization on the long pole) and set input_precision to ap_fixed<12,2> to match the deployed quantization. Co-authored-by: Cursor <cursoragent@cursor.com>
Point SI_Toolkit at the Development commit adding the positional weight-loading fallback required to load the Dense long-pole net. Co-authored-by: Cursor <cursoragent@cursor.com>
Restore m_cart=0.230, u_max=1.77, M_fric=3.22 -- the only set confirmed to swing up and balance on the physical cartpole with RPGD. The honest decorrelated system-ID values (m_cart=2.82, u_max~21.5, M_fric~36) and why the 13/22 set broke RPGD are documented in Driver/DataAnalysis/MotorAndCartFriction/motor_force_identification_2026-06/README.md. Co-authored-by: Cursor <cursoragent@cursor.com>
This is the long-pole Dense net used by the PC neural-imitator controller (swing-up + stabilization). It was only on disk (the Experiments/ folder is gitignored), so force-add it to make the working MLP reproducible from a clone. Its normalization matches the firmware HLS4ML 'v1' block (x3232_12_2_v1). Co-authored-by: Cursor <cursoragent@cursor.com>
- config_controllers.yml: neural-imitator points to the committed Dense-7IN-32H1-32H2-1OUT-8 net, nn_evaluator_mode 'C' (same math as the on-chip pure-C controller); keep the paper/FPGA 1OUT-1 net as a documented alternative. - Convert_Network_To_C.py: target 1OUT-8 for C-code generation. - SI_Toolkit: pull in network_parameters.c compile fix. Co-authored-by: Cursor <cursoragent@cursor.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Note
High Risk
High risk because it changes core CartPole dynamics/actuation plumbing (new neural-model stepping and acceleration modes, updated
Q2u, renamed noise parameters) and adds a new ZeroMQ-based remote control path, all of which can materially alter simulation fidelity and runtime behavior.Overview
Adds configurable dynamics execution paths to the simulator: CartPole can now advance state via
next_step_mode(Eulervs neural net) and compute accelerations viasecond_derivatives_mode(ODE, neural-only, or ODE+NN residual), with updatedQ2u(Q, u_max)usage throughout.Reworks control disturbance injection by replacing
add_control_noise/controlDisturbance*with a statefulControlNoiseGeneratorbacked by a new time-step-invariantCorrelatedNoiseGenerator, and renames exported parameters tocontrolNoiseScale/Bias/Correlation/controlNoise_mode.Introduces a new
cartpole_ekf.py(EKF + adaptive Q/R tuner with YAML persistence + live reload), updates CSV logging/header metadata, tweaks MPC cost/optimizer configs, and adds an additional neural controller (controller_difflg) plus bundled model artifacts.Modernizes the Gym env by decoupling physics from tasks and rendering: adds
Taskregistry (tasks.py), aPygameViewer, aCartPoleSimulatorBaseinterface, replaces the legacy sensors sim withCartpole_CustomSim, and addsCartpole_Remoteto run the environment against physical hardware over ZeroMQ.Written by Cursor Bugbot for commit 95c12c2. This will update automatically on new commits. Configure here.