Isaac Lab Integration

Safety middleware for NVIDIA Isaac Lab robot learning environments using THSP validation protocol.

Installation

pip install sentinelseed

# Isaac Lab (separate installation required)
# See: https://isaac-sim.github.io/IsaacLab/

Quick Start

from sentinelseed.integrations.isaac_lab import (
    SentinelSafetyWrapper,
    RobotConstraints,
)

# Create Isaac Lab environment
env = gym.make("Isaac-Reach-Franka-v0", cfg=cfg)

# Wrap with safety validation
env = SentinelSafetyWrapper(
    env,
    constraints=RobotConstraints.franka_default(),
    mode="clamp",  # clamp, block, warn, monitor
)

# Use normally - actions validated automatically
obs, reward, done, truncated, info = env.step(action)

Safety Modes

Robot Constraints

Pre-built

# Franka Emika Panda (7 DOF)
constraints = RobotConstraints.franka_default()

# Universal Robots UR10 (6 DOF)
constraints = RobotConstraints.ur10_default()

# Joint limits
limits = JointLimits.franka_panda()
limits = JointLimits.ur10()
limits = JointLimits.allegro_hand()

Custom

from sentinelseed.integrations.isaac_lab import (
    JointLimits,
    WorkspaceLimits,
    RobotConstraints,
    CollisionZone,
)

joint_limits = JointLimits(
    num_joints=6,
    position_lower=[-3.14, -2.0, -3.14, -3.14, -3.14, -3.14],
    position_upper=[3.14, 2.0, 3.14, 3.14, 3.14, 3.14],
    velocity_max=[2.0, 2.0, 2.0, 2.5, 2.5, 2.5],
)

workspace = WorkspaceLimits(
    x_min=-0.5, x_max=0.5,
    y_min=-0.5, y_max=0.5,
    z_min=0.0, z_max=1.0,
)

constraints = RobotConstraints(
    joint_limits=joint_limits,
    workspace_limits=workspace,
)

constraints.add_collision_zone(
    CollisionZone.sphere("human", center=(0.3, 0.0, 0.5), radius=0.2)
)

THSP Gates for Robotics

Truth Gate

• No NaN or Inf values
• Values within expected range
• Action dimensions match robot

Harm Gate

• Joint velocities within limits
• Forces/torques within safe ranges

Scope Gate

• Joint positions within mechanical limits
• End-effector within workspace
• No collision zone violations

Training Integration

Stable-Baselines3

from stable_baselines3 import PPO
from sentinelseed.integrations.isaac_lab import SentinelSB3Callback

env = SentinelSafetyWrapper(base_env, mode="clamp")
model = PPO("MlpPolicy", env)
callback = SentinelSB3Callback(env, log_interval=1000)

model.learn(total_timesteps=100000, callback=callback.get_sb3_callback())

Weights & Biases

from sentinelseed.integrations.isaac_lab import create_wandb_callback

callback = create_wandb_callback(
    env,
    project="my-robot-training",
    log_interval=100,
)

References

• Isaac Lab Documentation
• Gymnasium Wrappers
• Safe RL Survey