NED2 — Reach

The arm must move its end-effector to a 3D target sampled in the workspace above the cafe-table. No cube; gripper not commanded.

Env IDs: NED2ReacherSim-v0 / NED2ReacherGoalSim-v0 / NED2ReacherReal-v0 / NED2ReacherGoalReal-v0.

Description

A Niryo Ned2 6-DoF arm with 2 prismatic mors fingers sits on the same cafe-table the RX200 uses. URDF link names are bare (no ned2/ prefix on the URDF, although Gazebo’s link-state lookup uses the model-qualified ned2/<link> form). MoveIt SRDF group arm chains base_link → tool_link.

Action Space

Joint mode (default). Box(6,):

Num

Action

Min

Max

Joint

Unit

0

joint_1 delta (shoulder rot)

-2.949

+2.949

joint_1

rad

1

joint_2 delta (arm rot)

-1.833

+0.610

joint_2

rad

2

joint_3 delta (elbow rot)

-1.340

+1.570

joint_3

rad

3

joint_4 delta (forearm rot)

-2.090

+2.090

joint_4

rad

4

joint_5 delta (wrist rot)

-1.920

+1.923

joint_5

rad

5

joint_6 delta (hand rot)

-2.530

+2.530

joint_6

rad

Joint-1 bounds tightened from URDF’s ±3.000 to ±2.949 as a 0.05 rad safety margin. delta_action=True (default) scales by delta_coeff = 0.05.

EE mode (ee_action_type=True). Box(3,) — Δ EE position relative to base_link.

Observation Space

Standard env. Box layout:

  • EE position (3, base_link frame, m)

  • Unit vector EE → goal (3, normalised)

  • Distance EE → goal (1, m)

  • Current joint positions (8 — alphabetical: joint_1, joint_2, joint_3, joint_4, joint_5, joint_6, mors_1, mors_2)

  • Previous action (6 or 3)

  • Current joint velocities (8)

Goal env. Dict with three keys. desired_goal / achieved_goal = Box(3,). The bounds below are the declared observation-space bounds (mirror position_desired_goal_min/max in ned2_reach_task_config.yaml); the per-episode sampling support is the narrower position_goal_min/max (see Goal sampling below).

Idx

Dim

Component

Min

Max

0

1

goal x

0.15

0.40

1

1

goal y

-0.40

0.40

2

1

goal z

0.10

0.40

Rewards

Sparse: 0.0 if ‖ee − goal‖ < 0.02 else -1.0.

Dense: same shape as the other robots’ reach envs. Defaults from config/ned2_reach_task_config.yaml: reach_tolerance=0.02, multiplier_dist_reward=2.0, reached_goal_reward=20, step_reward=-0.5.

env = gym.make("NED2ReacherSim-v0", reward_type="Dense")
env = gym.make("NED2ReacherGoalSim-v0", reward_type="Sparse")

Starting State

Initial joint pose: URDF zero (init_pos = [0.0, 0.0, 0.0, 0.0, 0.0, 0.0] — Niryo’s “fully extended forward and up” pose; consider switching to a Niryo-style sleep pose like [0.0, 0.55, -1.5, 0.0, 0.0, 0.0] if your bench lacks vertical clearance).

Goal sampling. desired_goal ∈ Box(3,) from position_goal_min/max. Per-link FK safety enforces the table_z + safety_z_margin floor.

Episode End

Truncation. max_episode_steps (default 100). Real env aborts on stale /ned2/joint_states.

Termination. ‖ee − goal‖ < reach_tolerance (sparse only).

Arguments

Same kwargs as UR5e — Reach, plus NED2-specific:

  • use_wrist_camera (bool, default False) — subscribes to /gazebo_camera/image_raw (sim) or /niryo_robot_vision/compressed_video_stream (real) for Niryo’s built-in tool camera. Decoded frame exposed as self.cv_image_wrist.

Version History

  • v0 — first release (rl_environments v0.1.0). Bare URDF link names. MoveIt SRDF group arm ends at tool_link.