Finger

A two-DoF planar finger interacting with a free-rotating spinner body. The family covers two task styles: continuously spinning the body, and rotating it to a target angle.

FingerSpin

Property	Value
Canonical ID	mjx/finger_spin-v0
Action space	Box(-1.0, 1.0, (2,), float32)
Observation space	Box(-inf, inf, (9,), float32)
Episode length	1000
Config	{"ctrl_dt": 0.02, "sim_dt": 0.005, "naconmax": 25_000, "njmax": 50}

Description

The finger has to keep the spinner rotating around its hinge. There's no target angle — only the spinner's instantaneous angular velocity matters. The challenge is producing a coordinated finger motion that builds and sustains spinner momentum without snagging.

Rewards

Uses a sparse reward that fires only when the spinner is rotating fast enough in the negative direction:

reward = (hinge_velocity <= -SPIN_VELOCITY).astype(float)

The threshold gives a clean binary signal:

• 1.0 when the spinner is rotating at or faster than SPIN_VELOCITY in the negative direction.
• 0.0 otherwise.

Starting state

obs = [-1.5711 -0.5291  0.0702 -0.1094  0.      0.      0.      0.      0.    ]

(finger joint positions, spinner orientation, and joint velocities — finger initialised at a default configuration with the spinner at rest.)

Termination

Episode ends when step >= max_steps (default 1000). No early termination.

Usage

import envrax
env = envrax.make("mjx/finger_spin-v0")

Reference

Upstream: mujoco_playground/_src/dm_control_suite/finger.py.

---

FingerTurnEasy

Property	Value
Canonical ID	mjx/finger_turn_easy-v0
Action space	Box(-1.0, 1.0, (2,), float32)
Observation space	Box(-inf, inf, (12,), float32)
Episode length	1000
Config	{"ctrl_dt": 0.02, "sim_dt": 0.005, "naconmax": 25_000, "njmax": 50}

Description

The finger has to rotate the spinner to a randomised target angle and hold it there. The "easy" variant uses a generous EASY_TARGET_SIZE tolerance around the target, so the agent has a wide acceptance band to find — useful for bootstrapping a controller before tightening the criterion in FingerTurnHard.

Rewards

Uses a sparse reward that fires once the spinner has rotated to within tolerance of the target angle:

reward = (dist_to_target <= 0.0).astype(float)
# dist_to_target uses the EASY_TARGET_SIZE radius

dist_to_target is signed — negative inside the band, positive outside — so the comparison gives a clean binary signal:

• 1.0 when the spinner sits within EASY_TARGET_SIZE of the target angle.
• 0.0 otherwise.

Starting state

obs = [-1.5711 -0.5291  0.0702 -0.1094  0.      0.      0.      0.      0.
       -0.2     0.      0.2215]

(finger and spinner state plus the target's position fields appended at the end.)

Termination

Episode ends when step >= max_steps (default 1000). No early termination.

Usage

import envrax
env = envrax.make("mjx/finger_turn_easy-v0")

Reference

Upstream: mujoco_playground/_src/dm_control_suite/finger.py.

---

FingerTurnHard

Property	Value
Canonical ID	mjx/finger_turn_hard-v0
Action space	Box(-1.0, 1.0, (2,), float32)
Observation space	Box(-inf, inf, (12,), float32)
Episode length	1000
Config	{"ctrl_dt": 0.02, "sim_dt": 0.005, "naconmax": 25_000, "njmax": 50}

Description

The same target-rotation task as FingerTurnEasy, but the tolerance shrinks to HARD_TARGET_SIZE. With the smaller acceptance band, random exploration almost never lands inside the target — the variant doubles as a sparse-exploration benchmark on top of the manipulation task itself.

Rewards

Uses a sparse reward that fires once the spinner has rotated to within the tighter tolerance band:

reward = (dist_to_target <= 0.0).astype(float)
# dist_to_target uses the HARD_TARGET_SIZE radius

dist_to_target is signed — negative inside the band, positive outside — so the comparison gives a clean binary signal:

• 1.0 when the spinner sits within HARD_TARGET_SIZE of the target angle.
• 0.0 otherwise.

Starting state

obs = [-1.5711 -0.5291  0.0702 -0.1094  0.      0.      0.      0.      0.
       -0.2     0.      0.2615]

Termination

Episode ends when step >= max_steps (default 1000). No early termination.

Usage

import envrax
env = envrax.make("mjx/finger_turn_hard-v0")

Reference

Upstream: mujoco_playground/_src/dm_control_suite/finger.py.