Are there any reinforcement learning toolkits that handle sample-efficient locomotion training with safety constraints?

Yes. There are several RL toolkits and research frameworks that target sample-efficient locomotion and/or safety-constrained training, though it’s still an active research area and no single toolkit is perfect at both.

Toolkits/frameworks worth looking at

• Stable-Baselines3 + custom safety layer

  • Very popular and easy to use.
  • Good for prototyping locomotion.
  • Not safety-focused out of the box, but you can add:
    • action shielding
    • reward penalties
    • constraint checks
    • curriculum learning
  • Best if you want flexibility rather than built-in safety guarantees.

• Ray RLlib

  • Scales well and supports distributed training.
  • Useful for sample efficiency via parallelism.
  • Can be extended with custom constraint logic.
  • More engineering-heavy, but good for larger experiments.

• OmniSafe

  • A toolkit specifically for safe reinforcement learning.
  • Includes constrained RL algorithms such as Lagrangian-style methods and safe policy optimization variants.
  • More aligned with safety constraints than most general-purpose libraries.

• Safety Gymnasium / Safety-Gym

  • Benchmarks and environments designed for constrained/safe RL.
  • Often used with locomotion-like tasks and safety cost signals.
  • Not a full training library by itself, but pairs well with safe RL algorithms.

• Isaac Lab / Isaac Gym

  • Excellent for sample-efficient locomotion due to GPU-accelerated simulation.
  • Often used for quadruped and humanoid locomotion.
  • Safety constraints are not the main focus, but you can implement them in the environment or policy wrapper.
  • Great if sample efficiency is a priority.

• Brax

  • Very fast physics simulation on accelerator hardware.
  • Strong for sample-efficient locomotion training.
  • Like Isaac, it’s not primarily a safety toolkit, but it can be combined with constrained methods.

• RL Games

  • Frequently used with Isaac Gym for locomotion tasks.
  • Very sample-efficient in practice.
  • Safety constraints need to be added externally.

If you want both sample efficiency and safety

A common practical stack is:

• fast simulator: Isaac Gym / Isaac Lab or Brax
• safe RL algorithm: from OmniSafe or a constrained-optimization implementation
• environment cost signals: joint limits, torque limits, falls, self-collision, terrain hazards
• safety wrapper/shield: hard constraints at action or environment level

Rule of thumb

• If you want best sample efficiency: Isaac Lab/Gym or Brax
• If you want built-in safety constraints: OmniSafe or Safety Gymnasium
• If you want both: combine a fast simulator with a safe RL library and custom constraints

If you want, I can also give you:

1. a short list of the best options for quadrupeds/humanoids, or
2. a recommended stack for your specific robot and simulator.