CT in Gazebo

[kaiyun717]

Old Danaus Training

param_dict = {
        "m": 0.67634104,
        "J_xx": 0.00320868,
        "J_xy": 0.00011707,
        "J_xz": 0.00004899,
        "J_yy": 0.00288707,
        "J_yz": 0.00006456,
        "J_zz": 0.00495141,
        "angle": 0.9222,    # 52.84 degrees (rotor arm to x-axis)
        "r1": 0.11053858,   # Rotor arm length for rotors 1 (FR) and 3 (RL)
        "r2": 0.11232195,   # Rotor arm length for rotors 2 (FL) and 4 (RR)
        "m_s": 0.0183,      # Moment scale for 2204-2300KV motors
        "m_p": 0.03133884,      # Mass of pendulum
        "L_p": 0.5*2,       # This is the LENGTH OF PENDULUM. Total length: 1.085m
        "max_thrust": 4,    # Max thrust for each motor - set to 1200g at 100% throttle
        "min_thrust": 0.00, # Min thrust for each motor - set to 0g at 0% throttle
        'delta_safety_limit': math.pi / 4  # should be <= math.pi/4
    }

Commands

old_1

python main.py --reg_transform "sigmoid" --reg_sampler "random_inside" --reg_weight 0.0 --objective_option "weighted_average" --phi_nnl "tanh-tanh-softplus" --phi_nn_inputs "euc" --phi_include_xe --train_attacker_n_samples 50 --train_attacker_use_n_step_schedule --train_attacker_max_n_steps 20 --train_attacker_p_reuse 0.0 --train_attacker gradient_batch_warmstart_faster --gradient_batch_warmstart_faster_speedup_method sequential --gradient_batch_warmstart_faster_sampling_method gaussian --random_seed 0 --trainer_n_steps 1500 --affix "old_p_reuse_0_bs_50_reg_0" --gpu 0

old_2

python main.py --reg_transform "sigmoid" --reg_sampler "random_inside" --reg_weight 200.0 --objective_option "weighted_average" --phi_nnl "tanh-tanh-softplus" --phi_nn_inputs "euc" --phi_include_xe --train_attacker_n_samples 50 --train_attacker_use_n_step_schedule --train_attacker_max_n_steps 20 --train_attacker_p_reuse 0.0 --train_attacker gradient_batch_warmstart_faster --gradient_batch_warmstart_faster_speedup_method sequential --gradient_batch_warmstart_faster_sampling_method gaussian --random_seed 0 --trainer_n_steps 1500 --affix "old_p_reuse_0_bs_50_reg_200" --gpu 1

old_3

python main.py --reg_transform "sigmoid" --reg_sampler "random_inside" --reg_weight 0.0 --objective_option "weighted_average" --phi_nnl "tanh-tanh-softplus" --phi_nn_inputs "euc" --phi_include_xe --train_attacker_n_samples 100 --train_attacker_use_n_step_schedule --train_attacker_max_n_steps 20 --train_attacker_p_reuse 0.0 --train_attacker gradient_batch_warmstart_faster --gradient_batch_warmstart_faster_speedup_method sequential --gradient_batch_warmstart_faster_sampling_method gaussian --random_seed 0 --affix "old_critic_bs_100_reg_0" --gpu 2

old_4

python main.py --reg_transform "sigmoid" --reg_sampler "random_inside" --reg_weight 200.0 --objective_option "weighted_average" --phi_nnl "tanh-tanh-softplus" --phi_nn_inputs "euc" --phi_include_xe --train_attacker_n_samples 100 --train_attacker_use_n_step_schedule --train_attacker_max_n_steps 20 --train_attacker_p_reuse 0.5 --train_attacker gradient_batch_warmstart_faster --gradient_batch_warmstart_faster_speedup_method sequential --gradient_batch_warmstart_faster_sampling_method gaussian --random_seed 0 --affix "old_critic_bs_100_reg_200" --gpu 3

noise_old_1

python main.py --reg_transform "sigmoid" --reg_sampler "random_inside" --reg_weight 0.0 --objective_option "weighted_average" --phi_nnl "tanh-tanh-softplus" --phi_nn_inputs "euc" --phi_include_xe --train_attacker_n_samples 50 --train_attacker_use_n_step_schedule --train_attacker_max_n_steps 20 --train_attacker_p_reuse 0.0 --train_attacker gradient_batch_warmstart_faster --gradient_batch_warmstart_faster_speedup_method sequential --gradient_batch_warmstart_faster_sampling_method gaussian --random_seed 0 --trainer_n_steps 1500 --state_noise 0.01 --affix "pend_noise_old_p_reuse_0_bs_50_reg_0" --gpu 0

noise_old_2

python main.py --reg_transform "sigmoid" --reg_sampler "random_inside" --reg_weight 200.0 --objective_option "weighted_average" --phi_nnl "tanh-tanh-softplus" --phi_nn_inputs "euc" --phi_include_xe --train_attacker_n_samples 50 --train_attacker_use_n_step_schedule --train_attacker_max_n_steps 20 --train_attacker_p_reuse 0.0 --train_attacker gradient_batch_warmstart_faster --gradient_batch_warmstart_faster_speedup_method sequential --gradient_batch_warmstart_faster_sampling_method gaussian --random_seed 0 --trainer_n_steps 1500 --state_noise 0.01 --affix "pend_noise_old_p_reuse_0_bs_50_reg_200" --gpu 1

noise_old_3

python main.py --reg_transform "sigmoid" --reg_sampler "random_inside" --reg_weight 0.0 --objective_option "weighted_average" --phi_nnl "tanh-tanh-softplus" --phi_nn_inputs "euc" --phi_include_xe --train_attacker_n_samples 100 --train_attacker_use_n_step_schedule --train_attacker_max_n_steps 20 --train_attacker_p_reuse 0.0 --train_attacker gradient_batch_warmstart_faster --gradient_batch_warmstart_faster_speedup_method sequential --gradient_batch_warmstart_faster_sampling_method gaussian --random_seed 0 --trainer_n_steps 1500 --state_noise 0.01 --affix "pend_noise_old_critic_bs_100_reg_0" --gpu 2

noise_old_4

python main.py --reg_transform "sigmoid" --reg_sampler "random_inside" --reg_weight 200.0 --objective_option "weighted_average" --phi_nnl "tanh-tanh-softplus" --phi_nn_inputs "euc" --phi_include_xe --train_attacker_n_samples 100 --train_attacker_use_n_step_schedule --train_attacker_max_n_steps 20 --train_attacker_p_reuse 0.5 --train_attacker gradient_batch_warmstart_faster --gradient_batch_warmstart_faster_speedup_method sequential --gradient_batch_warmstart_faster_sampling_method gaussian --random_seed 0 --trainer_n_steps 1500 --state_noise 0.01 --affix "pend_noise_old_critic_bs_100_reg_200" --gpu 3

[kaiyun717]

Notes

• $\omega_z$ is set to be 0 regardless of the QP output to ensure that we don't do yaw-rate control at all.
• Gazebo is running at 0.5 real-time.

Checklists

• [ ] Verify ROS-Gazebo time synchronizer
• [ ] See what the effect of changing the RK4 frequency (dt) is
• [ ] Check which ckpt + training has the largest invariant sizes
• [ ] If nothing works, try adding $-\alpha(\phi(x))$ (?).