Implement GIGAFLOW dynamics model (jerk-based control) #1

elmashadmo · 2025-08-29T15:04:58Z

Add option to switch between simple acceleration controls and jerk-based physics
Change action space from [7,13] to [4,3] for longitudinal/lateral jerk when in GIGAFLOW
Add new observation features for current acceleration and steering
Include comprehensive test suite with 9 test cases for math verification
Maintain backward compatibility with existing classic dynamics
Include 4 maps
Update .gitignore

eugenevinitsky · 2025-08-31T13:55:41Z

pufferlib/ocean/drive/drive.h

+    // Clamp indexes
+    // if (long_jerk_idx < 0) long_jerk_idx = 0;
+    // if (long_jerk_idx > 3) long_jerk_idx = 3;
+
+    // if (lat_jerk_idx < 0) lat_jerk_idx = 0;
+    // if (lat_jerk_idx > 2) lat_jerk_idx = 2;


should this be commented out?

Yes, this was addressing a bug I had but I ended up fixing it

eugenevinitsky · 2025-08-31T13:56:15Z

pufferlib/ocean/drive/drive.h

+        // TODO Need to test this in environment
+        // Rotate to world frame
+        // float temp_x = dx;
+        // dx = temp_x * cosf(agent->heading) - dy * sinf(agent->heading);
+        // dy = temp_x * sinf(agent->heading) + dy * cosf(agent->heading);
+    }


same here, unsure if PR is ready

eugenevinitsky · 2025-08-31T13:58:04Z

pufferlib/ocean/drive/drive.h

+    // 3 additional state variables per agent in GIGAFLOW
+    int max_obs = 7 + 3 + 7*(MAX_CARS - 1) + 7*MAX_ROAD_SEGMENT_OBSERVATIONS;


shouldn't this size be conditional on the dynamics model?

eugenevinitsky · 2025-08-31T13:58:49Z

pufferlib/ocean/drive/drive.h

+
        // Relative Pos of other cars
-        int obs_idx = 7;  // Start after goal distances
+        int obs_idx = 9;  // Start after goal distances + GIGAFLOW additions


should this be +2 or +3?

eugenevinitsky · 2025-09-01T14:53:00Z

.gitignore

+
+# macOS Build Artifacts
+*.dSYM/
+test_gigaflow_dynamics


is this the correct path?

* Add option to switch between simple acceleration controls and jerk-based physics * Change action space from [7,13] to [4,3] for longitudinal/lateral jerk when in GIGAFLOW * Add new observation features for current acceleration and steering * Include comprehensive test suite with 9 test cases for math verification * Maintain backward compatibility with existing classic dynamics * Include 4 maps * Update .gitignore

…rated with observations, collision, grid_map

nadarenator · 2025-09-09T17:21:21Z

pufferlib/ocean/drive/binding.c

    sprintf(map_file, "resources/drive/binaries/map_%03d.bin", map_id);
    env->num_agents = max_agents;
    env->map_name = strdup(map_file);
+    env->dt = 0.1f;


Would be nice if this could be passed as an env arg

nadarenator · 2025-09-09T17:25:08Z

pufferlib/ocean/drive/drive.h

+    float a_lat;       // Current lateral acceleration
+    float prev_a_lat;  // Previous lateral acceleration
+
+    float phi;         // Current steering angle


I think renaming this to steering_angle would be more readable

nadarenator

Summary of requested changes:

remove classic dynamics
pass dt as an env arg
rename gigaflow dynamics to something more descriptive
rename phi
remove randomization coefficients and sampling
rename obs_size to ego_obs_size or self_obs_size
single move_dynamics function with branching
remove TODO on test_performance function if done

nadarenator · 2025-09-09T17:31:28Z

pufferlib/ocean/drive/drive.h

+    float c_throttle;  // Throttle responsiveness
+    float c_steer;     // Steering responsiveness
+    float c_acc;       // Acceleration limits multiplier
+    float c_vel;       // Velocity limits multiplier


I don't think randomization coefficients are necessary in the main branch, can be scoped to a specific project. Anything that would be generally useful across projects should be in the main branch imo.

Is the reasoning for removing it due to it being more impactful to how the agents train and less impactful to the realism/dynamics portion? I removed it though for the time being!

You're right yeah. To give you some more context, in my opinion the main branch should be minimal, with features that would be useful across different projects. The jerk-based dynamics is a good example of such a feature. But dynamics randomization or even behavior conditioning coefficients is specific only to training a population of agents, which a lot of projects may not require.

nadarenator · 2025-09-09T17:39:32Z

pufferlib/ocean/drive/drive.h

+    entity->prev_phi = 0;
+    entity->prev_v = 0;
+}
+


Like the previous comment, I think random sampling and coefficients should be scoped to a project and not main

nadarenator · 2025-09-09T17:43:37Z

pufferlib/ocean/drive/drive.h

        e->collision_state = 0;
        e->respawn_timestep = -1;
+          // Initialize GIGAFLOW stateful variables if using GIGAFLOW dynamics
+        if (env->dynamics_model == GIGAFLOW && is_active) {


For people who haven't read the paper, GigaFlow doesn't really mean anything. A more informative name for dynamics could be jerk, jerk_bicycle or jerk_curvature.

nadarenator · 2025-09-09T17:45:50Z

pufferlib/ocean/drive/drive.h

    init(env);
-    int max_obs = 7 + 7*(MAX_CARS - 1) + 7*MAX_ROAD_SEGMENT_OBSERVATIONS;
+    // 3 additional state variables per agent in GIGAFLOW
+    int obs_size = (env->dynamics_model == GIGAFLOW) ? 10 : 7;


Could you rename this to self_obs_size or ego_obs_size?

nadarenator · 2025-09-09T17:48:32Z

pufferlib/ocean/drive/drive.h

    return heading;
 }

+void move_gigaflow_dynamics(Drive* env, int action_idx, int agent_idx){


Would be better if we stick to a single move_dynamics function that branches.

nadarenator · 2025-09-09T18:12:42Z

pufferlib/ocean/drive/drive.py

-        self.single_observation_space = gymnasium.spaces.Box(low=-1, high=1,
+        self.dynamics_model = dynamics_model
+
+        if dynamics_model == "classic":


I don't think we need classic dynamics at all in the env, since the gigaflow dynamics is better all-round.

nadarenator · 2025-09-09T18:14:34Z

pufferlib/ocean/drive/drive.py

        # except Exception as e:
        #     print(f"Error processing {map_path.name}: {e}")

+# TODO: Update test_performance


Is this done?

…ent-road Split entities to agent and road

This commit addresses changes in the environment dynamics for improved clarity, control, and simplicity. * Passes dt as an environment argument for more flexible simulation * Removes randomization coefficients and related sampling logic * Improves variable naming for better readability (gigaflow_dynamics, phi, obs_size) * Removes a completed TODO in the test_performance function

…ynamics

eugenevinitsky · 2025-10-20T16:47:19Z

Closing because duplicated with another PR. Props all the same @elmashadmo

Pragnay Mandavilli added 2 commits August 13, 2025 10:19

Added profiling scripts

a4ffa79

gitignore changes

27f9f39

eugenevinitsky reviewed Aug 31, 2025

View reviewed changes

eugenevinitsky reviewed Sep 1, 2025

View reviewed changes

.gitignore Outdated

# macOS Build Artifacts

*.dSYM/

test_gigaflow_dynamics

Copy link

eugenevinitsky Sep 1, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

is this the correct path?

elmashadmo force-pushed the gigaflow-dynamics branch from 255244c to c3d135e Compare September 3, 2025 03:31

Pragnay Mandavilli and others added 4 commits September 4, 2025 12:59

Added Env SPS, Eval SPS and ram_tracking logs

2af0dfd

Code Cleanup

720c255

Split Entities Struct to MapEntities, added enums for types and integ…

b9a0d0c

…rated with observations, collision, grid_map

Fixed free_entities for separate types, some code cleanup warnings

b7ded52

nadarenator reviewed Sep 9, 2025

View reviewed changes

nadarenator requested changes Sep 9, 2025

View reviewed changes

Pragnay Mandavilli and others added 3 commits September 9, 2025 16:50

Removed debug prints, configs

826e468

Merge pull request Emerge-Lab#25 from Emerge-Lab/split-entities-to-ag…

17e85f5

…ent-road Split entities to agent and road

elmashadmo force-pushed the gigaflow-dynamics branch from d2770f5 to 241ae49 Compare September 17, 2025 08:35

Merge branch '3.0' of github.com:Emerge-Lab/PufferLib into gigaflow-d…

5128819

…ynamics

nadarenator changed the base branch from 3.0 to main September 23, 2025 19:16

nadarenator added 2 commits September 24, 2025 11:50

merged latest main

2c6512e

revert gitignore

7858a74

eugenevinitsky closed this Oct 20, 2025

daphne-cornelisse mentioned this pull request Nov 24, 2025

Add mode for controlled experiments #144

Merged

		// 3 additional state variables per agent in GIGAFLOW
		int max_obs = 7 + 3 + 7(MAX_CARS - 1) + 7MAX_ROAD_SEGMENT_OBSERVATIONS;

Implement GIGAFLOW dynamics model (jerk-based control) #1

Implement GIGAFLOW dynamics model (jerk-based control) #1

Uh oh!

Conversation

elmashadmo commented Aug 29, 2025

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

nadarenator Sep 9, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

nadarenator left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

eugenevinitsky commented Oct 20, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

nadarenator Sep 9, 2025 •

edited

Loading