IROS 2024 Pendubot Results

Controller	Short Controller Description	Swingup Success	Swingup Time [s]	Energy [J]	Torque Cost [N²m²]	Torque Smoothness [Nm]	Velocity Cost [m²/s²]	Best RealAI Score	Average RealAI Score	Username	Data
ar_eapo	Policy trained with average reward maximum entropy RL	10/10	0.77	11.7	7.33	0.237	49.53	0.711	0.646	rnilva	data plot video
MC-PILCO	Controller trained with the Model-Based Reinforcement Learning algorithm MC-PILCO for swing-up and stabilization around the unstable equilibrium of the Pendubot.	10/10	0.73	10.22	9.92	0.475	36.84	0.691	0.644	turcato-niccolo	data plot video
history_sac	SAC using custom model architecture to encode system dynamics.	7/10	5.73	13.98	5.08	0.197	141.48	0.509	0.343	tfaust	data plot video
evolsac	Evolutionary SAC for both swingup and stabilisation	2/10	5.02	23.16	18.59	0.346	243.33	0.351	0.068	AlbertoSinigaglia	data plot video

Videos from left to right: AR_EAPO, MC-PILCO, History_SAC, EvolSAC

Rules

This leaderboard shows the final results from the RealAIGym competition at IROS 2024.

The real system leaderboard compares the performance of different control methods on the real hardware. The task for the controller is to swingup and balance the pendubot and keep the end-effector above the threshold line.

During the execution external perturbations in the form of Gaußian torque peaks have been applied by the motors at both links. In total there were 4 perturbations, 2 at each joint, with a standard deviation between 0.05s and 0.1s and an amplitude between 0.2Nm and 0.5Nm. The perturbations were randomly generated and different in all 10 trials. The perturbations are the same for all controllers.

The model parameters identified by us with a least squares optimization of the pendubot are:

Gravity: \(g = 9.81 \, \frac{\text{m}}{\text{s}^2}\)
First Link mass: \(m_1 = 0.523 \, \text{kg}\)
Second Link mass: \(m_2 = 0.625 \, \text{kg}\)
First actuator length: \(l_1 = 0.2 \, \text{m}\)
Second actuator length: \(l_2 = 0.3 \, \text{m}\)
First link center of mass: \(r_1 = 0.2 \, \text{m}\)
Second link center of mass: \(r_2: 0.255 \, \text{m}\)
First link inertia: \(I_1 = 0.031 \, \text{kg m}^2\)
Second link inertia: \(I_2 = 0.050 \, \text{kg m}^2\)
First actuator damping friction: \(b_1 = 0.001 \, \frac{\text{kg m}}{s}\)
Second actuator damping friction: \(b_2 = 0.001 \, \frac{\text{kg m}}{s}\)
First actuator coulomb friction: \(c_{f1}: 0.16 \, \text{N m}\)
Second actuator coulomb friction: \(c_{f2}: 0.12 \, \text{N m}\)
Motor Inertia: \(I_r = 6.28e-05 \, \text{kg m}^2\)
Gear ratio: \(g_r: 6.0\)
First Motor Torque limit \(\tau_{max1} = 0.0 \, \text{N m}\)
Second Motor Torque limit \(\tau_{max2} = 6.0 \, \text{N m}\)

More information about the dynamic model of the double pendulum can be found here: Double Pendulum Dynamics. For a urdf file with this model see here: URDF.

The \(0.5\,\text{Nm}\) torque limit on the passive joint can be used to compensate the friction of the motor.

The actuators can be controlled with arbitrary control frequency of up to \(500\, \text{Hz}\) and the experiment takes \(10\,\text{s}\). The initial pendubot configuration is \(x_0 = (0, 0, 0, 0)\) (hanging down) and the goal is the unstable fixpoint at the upright configuration \(x_g = (\pi, 0, 0, 0)\). The upright position is considered to be reached when the end-effector is above the threshold line at \(h=0.45 \, \text{m}\) (origin at the mounting point).

Scores

For the evaluation multiple criteria are evaluated and weighted to calculate an overall score (Real AI Score). The criteria are:

Swingup Success \(c_{success}\): Whether the swingup was successful, i.e. if the end-effector is above the threshold line in the end of the simulation.
Swingup time \(c_{time}\): The time it takes for the pendubot to reach the goal region above the threhhold line and stay there. If the end-effector enters the goal region but falls below the line before the simulation time is over the swingup is not considered successful! The swingup time is the time when the end-effector enters the goal region and does not leave the region until the end.
Energy \(c_{energy}\): The mechanical energy used during the execution.
Torque Cost \(c_{\tau, cost}\): A quadratic cost on the used torques ( \(c_{\tau, cost} = \sum u^TRu\), with R = 1).
Torque Smoothness \(c_{\tau, smooth}\): The standard deviation of the changes in the torque signal.
Velocity Cost \(c_{vel, cost}\): A quadratic cost on the joint velocities that were reached during the execution ( \(c_{vel} = \dot{q}^T Q \dot{q}\), with Q = identity)

These criteria are used to calculate the overall Real AI Score with the formula

\[ \begin{equation} S = c_{success} \left( 1 - \sum_{i \in \{ 'time', 'energy', '\tau, cost', '\tau, smooth', 'vel, cost' \}} \tanh \left(\frac{c_{i}}{n_{i}}\right)\right) \end{equation} \]

The weights and normalizations are:

Criterion	normalization \(n\)
Swingup Time	20.0
Energy	60.0
Torque Cost	100
Torque Smoothness	4.0
Velocity Cost	400.0

The listed number for swingup time, energy, etc. in the leaderboard are the scores for from the best attempt. The ‘Best RealAIScore’ is the score of that attempt. The ‘Average RealAIScore’ is the average score over 10 attempts. Unsuccessful swingups (where the end effector is not above the threshhold line in the end of the experiment (i.e. after \(10\,\text{s}\))) have a score of 0.

Participating

This leaderboard is only for the results from the competition at IROS 2024. For participating checkout the ongoing leaderboard