Fxxkrobotics/rough_tactile

robosuite (trimmed handoff build)

A simulation setup based on robosuite (v1.5.2) for running robot manipulation tasks in MuJoCo and prototyping ideas.

The original tactile data-collection / policy-training (ACT) scaffolding has been removed. This copy keeps only the runnable simulation core so you can get started quickly.

---

What's in here

• Standard robosuite: built-in tasks (Lift, Stack, NutAssembly, PickPlace, Door, Wipe, two-arm, ...), controllers (OSC / JOINT / ...), and the usual robots and grippers.
• Tactile taxels on the gripper (with read-out): the Robotiq-85 gripper XML has 32 taxel_* sites (two 4×4 grids, one per fingerpad, group 4), plus wrist force/torque sensors force_ee / torque_ee. The reader uskin_sensor.py (repo root) maps MuJoCo contact forces onto those taxels — each taxel outputs a 3D force (fx, fy, fz), giving a 96-D observation. Meant for the Sawyer + Robotiq85 setup. See usage below.
• Extra task-scene assets under robosuite/models/assets/arenas/ (battery, book_shelf, usb, drawer, cap, ...) and matching object XMLs.
  • ⚠️ These XMLs are not referenced by any Python env in this copy (the env code that loaded them isn't included). Useful as building blocks for custom scenes, but not directly instantiable.

In short: this is standard robosuite plus some tactile / task-scene scaffolding assets. More than enough to run simulations and validate ideas.

---

Install

Requires Python ≥ 3.9, Linux + MuJoCo (pulled in by robosuite's mujoco dependency).

# recommended: a virtual environment
python -m venv .venv && source .venv/bin/activate

# editable install of this repo (installs robosuite and its dependencies)
pip install -e .

For simulation only the core deps are small — mainly mujoco, numpy, scipy. The requirements.txt at the repo root still lists some heavy training/collection deps (torch, wandb, h5py, ...) that are not needed to run simulations; ignore or remove them.

---

Quick start

# random actions in a task, with a render window
python -m robosuite.demos.demo_random_action

# keyboard / gamepad teleoperation
python -m robosuite.demos.demo_device_control

# controller demo (switch OSC / JOINT / ...)
python -m robosuite.demos.demo_control

More examples live in robosuite/demos/: multi-camera, domain randomization, segmentation, video recording, Gym interface, and so on.

Minimal custom script:

import robosuite as suite

env = suite.make(
    env_name="Lift",          # or Stack / NutAssembly / Door / PickPlace ...
    robots="Panda",           # or Sawyer / IIWA / UR5e ...
    has_renderer=True,
    has_offscreen_renderer=False,
    use_camera_obs=False,
)

obs = env.reset()
for _ in range(1000):
    action = env.action_spec[0] * 0  # replace with your policy output
    obs, reward, done, info = env.step(action)
    env.render()
env.close()

---

Tactile read-out (uSkin sensor)

uskin_sensor.py turns the gripper taxel sites into a tactile signal. Use it with the Sawyer + Robotiq85 setup:

import numpy as np
import robosuite as suite
from uskin_sensor import USkinSensor

env = suite.make(env_name="Lift", robots="Sawyer", gripper_types="Robotiq85Gripper",
                 has_renderer=False, has_offscreen_renderer=False, use_camera_obs=False)
env.reset()

sensor = USkinSensor(env.sim, gripper_prefix="gripper0_right_")
for _ in range(200):
    obs, r, done, info = env.step(np.zeros(env.action_spec[0].shape))
    reading = sensor.update()          # {"left_finger": (4,4,3), "right_finger": (4,4,3), ...}
    tactile_obs = sensor.get_observation()   # flat (96,) vector for a policy
env.close()

Readings are near-zero until the fingers actually make contact. Only depends on numpy and mujoco.

---

Docs & reference

• Official docs: https://robosuite.ai/docs/
• Env / controller / robot APIs under robosuite/environments/, robosuite/controllers/, robosuite/models/.

For real tactile observations or the extra scene tasks, you'd need to bring back the corresponding env code (not included in this trimmed copy).