Qualifying

In this stage, you will solve the three main subtasks with a 7dof KUKA IIWA robot. In the simulation, the collision between the robot and the table are disabled. Thus your solution must keep the mallet (Robot’s End Effector) at an appropriate height.

Note

During this phase, we will make different modifications in the environment to simulate real-world challenges. Participants will develop their robotic agents in an ideal environment. In order to deal with theses challenges, we encourage the participant to submit their agent, collect the dataset and improve the robustness of your agent.

Installation

1. If you did not join the warm-up stage, you can finish the Installation instruction.

2. If you have already cloned the repo, you can checkout to the qualifying branch and merge your local changes.

Environment Specifications

Here we list the some useful information about the environment.

Important

In the constraint, the joint position and velocity limits for constraint computation is 95% of the actual limits. For example, the upper bound of the position limit for joint 1 is 2.967. In the Evaluation and Constraints, we check if the joint position exceeds 2.967 * 0.95 = 2.818707.

Robot Specifications
Robot Position Upper Limit (rad)	[ 2.967, 2.09 , 2.967, 2.094, 2.967, 2.094, 3.054]
Robot Position Lower Limit (rad)	[-2.967, -2.094, -2.967, -2.094, -2.967, -2.094, -3.054]
Robot Velocity Limit (rad/s)	+/- [ 1.483, 1.483, 1.745, 1.308, 2.268, 2.356, 2.356]
Environment Specifications
Environments	7dof-hit, 7dof-defend, 7dof-prepare
Initial Robot Position (fixed)	[ 0, -0.1960, 0, -1.8436, 0, 0.9704 0]
Initial Robot Velocity	0

Hit: 7dof-hit

In this task, the opponent moves in a predictable pattern. The puck is initialized randomly with a small velocity. The task is to hit the puck to the opponent’s goal.

Initialization Range:

\(x\)	[0.81, 1.31]
\(y\)	[-0.39, 0.39]
linear speed (m/s)	[0, 0.5]
angular speed (rad/s)	[-2, 2]

Termination Criterion: The puck is bounced back or scored.

Success Criterion: The puck is in opponent’s goal when the episode terminates.

---

Defend: 7dof-defend

The puck is randomly initialized on the right side of the table with a random velocity heading left. The objective is to stop the puck on the right side of the table and prevent it from getting scored.

Initialization Range:

\(x\)	[1.80, 2.16]
\(y\)	[-0.4, 0.4]
linear speed (m/s)	[1, 3]
angular speed (rad/s)	[-10, 10]

Termination Criterion: The puck is returned to the opponent’s side or scored or the puck speed drops below the threshold.

Success Criterion: The puck is in the range where hits can be made and the longitudinal speed is below the threshold.

---

Prepare: 7dof-prepare

The puck is initialized close to the table’s boundary and is unsuitable for hitting. The task is to control the puck to move it into a good hit position. The puck is not allowed to cross the middle line.

Initialization Range:

position	[[0.71, 1.31], +/-[0.39105, 0.48535]] or
	[[0.57, 0.71], +/-[0.125, 0.48535]]
linear speed (m/s)	0.0
angular speed (rad/s)	0.0

Termination Criterion: Puck crosses the middle line that connects the middle points of two goals, or the puck is on the opponent’s side of the table.

Success Criterion: The puck is in the range where hits can be made and the longitudinal speed is below the threshold.

Action Interface

In this phase, we provide a flexible interface for commanding the robot. You can specify the trajectory interpolation_order order in the agent_config.yml. Here is the list of the interpolation:

3: Cubic interpolation. The action command contains the desired [position, velocity]. A cubic polynomial is used to interpolate the intermediate steps. The shape of the command should be [2, N_joints].

1: Linear interpolation. The action command contains the desired [position]. A linear polynomial is used to interpolate the intermediate steps. The shape of the command should be [N_joints]. Note that the acceleration is will be zero, the acceleration feedforward will also be zero.

2: Quadratic interpolation. The action command contains the desired [position]. A quadratic function uses the previous position, velocity and the desired position to interpolate the intermediate steps. The shape of the command should be [N_joints].

4: Quartic interpolation. The action command contains the desired [position, velocity]. A quartic function uses the previous position, velocity and the desired position, velocity to interpolate the intermediate steps. The shape of the command should be [2, N_joints].

5: Quintic interpolation. The action command contains the desired [position, velocity, acceleration]. A quintic function is computed by the previous position, velocity, acceleration and the desired position, velocity and acceleration to interpolate the intermediate steps. The shape of the command should be [3, N_joints].

-1: Linear interpolation in position and velocity. The action command contains the desired [position, velocity]. The position and velocity will both be linearly interpolated. The acceleration is computed based on the derivative of the velocity. This interpolation is not proper, but it is useful to avoid oscillatory in the interpolation. The shape of the command should be [2, N_joints].

None: You can send a complete trajectory between each action step. At each step, the trajectory command should include desired [position, velocity, acceleration]. The shape of the command should be [20, 3, N_joints].

Constraints

For 7 DoF Environments, additional constraints are added that ensures positions of the elbow and the wrist are above a threshold. The updated constraints table is listed here

Class Name	Key	Output Dim	Description
JointPositionConstraint	“joint_pos_constr”	2 * num_joints	\(q_l < q_{cmd} < q_u\)
JointVelocityConstraint	“joint_vel_constr”	2 * num_joints	\(\dot{q}_l < \dot{q}_{cmd} < \dot{q}_u\)
EndEffectorConstraint	“ee_constr”	5	\(l_x < x_{ee},\) \(l_y < y_{ee} < u_y,\) \(z_{ee} > \mathrm{table\,height - tolerance}\), \(z_{ee} < \mathrm{table\, height + tolerance}\).
LinkConstraint (7DoF Robot Only)	“link_constr”	2	\(z_{elbow} > 0.25\), \(z_{wrist} > 0.25\)

Evaluation

To evaluate your agent in the cloud server, please follow the Solution Submission instruction. The environments on the cloud server slightly differs to the public ones. It has additional challenges which occure in the real world. These challenge might be a model gap, error prone observations, etc.