This is the project related to the paper Slow but flexible or fast but rigid? Discrete and continuous processes compared. It investigates the tradeoff, in active inference, between high-level cognitive planning and low-level motor control regarding multi-step tasks. It compares two models (a hybrid discrete-continuous model and a continuous-only model) on a dynamic pick-and-place operation. The study further explores how discrete actions could lead to continuous attractors, and draws parallelisms with different motor learning phases, aiming to inform future research on bio-inspired task specialization.
Video simulations are found here.
Check this and this for additional projects and guides.
This study has received funding from the MAIA project, under the European Union's Horizon 2020 Research and Innovation programme.
The simulation can be launched through main.py, either with the option -m for manual control, -c for the continuous-only model, -y for the hybrid model, or -a for choosing the parameters from the console. If no option is specified, the continuous-only model will be launched. For the manual control simulation, the arm can be moved with the keys Z, X, LEFT, RIGHT, UP and DOWN. The hand can be rotated with A and S. The fingers can be open or closed for grasping with Q and W.
Plots can be generated through plot.py, either with the option -d for the hand-target distance and dynamics of action probabilities, -s for the scores, or -v for generating a video of the simulation.
More advanced parameters can be manually set from config.py. Custom log names are set with the variable log_name. The number of trials and steps can be set with the variables n_trials and n_steps, respectively.
The parameter grasp_dist controls the distance from the wrist at which the object is grasped.
The parameter open_angle controls the maximum opening of the hand.
The parameter n_tau specifies the sampling time window used for evidence accumulation.
The parameter target_vel specifies the velocity of the object, which in the simulations have been varied from 0 to 80 pixels per time step.
Different precisions and attractor gains have been used for the hybrid and continuous-only model: these are specified in main.py.
The agent configuration is defined through the dictionary joints. The value link specifies the joint to which the new one is attached; angle encodes the starting value of the joint; limit defines the min and max angle limits.
The scripts simulation/inference_cont.py and simulation/inference_hybrid.py contain subclasses of Window in environment/window.py, which is in turn a subclass of pyglet.window.Window. The only overriden function is update, which defines the instructions to run in a single cycle. Specifically, the subclasses InferenceContinuous and InferenceHybrid initialize the agent, the object and the goal; during each update, they retrieve proprioceptive, visual, and tactile observations through functions defined in environment/window.py, call the function inference_step of the respective agents (defined in cont_only_agent.py and hybrid_agent_cont.py), move the arm and the objects, and finally check for possible collisions. Every n_tau steps, InferenceHybrid calls an inference step of an additional discrete agent, defined in hybrid_agent_disc.py.
The continuous agents have a similar function inference_step, but with some differences. In particular, the continuous-only agent generates visual, proprioceptive, and tactile predictions via the function get_p, while the continuous agent of the hybrid method only generates visual and proprioceptive predictions. The continuous-only agent specifies transitions between intentions via the function set_intentions. This function generates variables gamma_int and gamma_ext, which affect the beliefs over hidden causes. Then, the function get_i computes the intentions from the intrinsic and extrinsic beliefs over hidden states and hidden causes. These are used to generate dynamics prediction errors via the function get_e_x. Instead, the continuous agent of the hybrid method generates prior prediction errors via the function get_e_x, depending on the full priors computed by the discrete agent. For both models, the function get_likelihood backpropagates the sensory prediction errors toward the beliefs, calling the function grad_ext to compute the extrinsic gradient. Finally, the function mu_dot computes the total belief updates. At each continuous time step, the continuous agent of the hybrid method accumulate evidence in both (intrinsic and extrinsic) modalities.
Bayesian model average and Bayesian model comparison between reduced priors is computed via the functions do_bmc and do_bma in the discrete agent of the hybrid method. The reduced priors are instead updated through the function get_eta_m.
Note that the function init_belief changes the length of the last joint to account for the grasping distance.
Useful trajectories computed during the simulations are stored through the class Log in environment/log.py.
Note that all the variables are normalized between -1 and 1 to ensure that every contribution to the belief updates has the same magnitude.
matplotlib==3.8.4
numpy==2.0.1
pyglet==2.0.17
pymunk==6.6.0
seaborn==0.13.2