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

Yes — there are a few RL toolkits and frameworks that are commonly used for sample-efficient locomotion and can support safety constraints, though the “safety” part is often more modular than built-in.

Good candidates

• Ray RLlib

  • Strong general-purpose RL framework.
  • Can be combined with custom safety logic, constrained optimization, or safe exploration methods.
  • Good for distributed training, but not specifically locomotion-focused.

• Stable-Baselines3 + safety extensions

  • Very popular for MuJoCo-style locomotion tasks.
  • Sample efficiency depends on the algorithm; off-policy methods like SAC/TD3 are often used.
  • Safety constraints usually require adding a wrapper or using a separate constrained/safe RL implementation.

• acme (DeepMind)

  • Flexible research framework.
  • Useful for sample-efficient algorithms and custom training loops.
  • Safety constraints are generally something you implement on top.

• Tianshou

  • Lightweight and flexible.
  • Supports off-policy algorithms suitable for sample efficiency.
  • Can be extended with custom safety critics, constraints, or shielding.

• Omniverse Isaac Gym / Isaac Lab

  • Particularly strong for locomotion because of fast parallel simulation.
  • Very sample-efficient in practice due to massive throughput.
  • Safety constraints can be integrated, but it’s not primarily a “safe RL” toolkit out of the box.

• CLeAR / safe RL research toolkits

  • There are research-oriented toolkits and implementations for:
    • Constrained Policy Optimization (CPO)
    • Lyapunov-based RL
    • Risk-sensitive / shielded RL
  • These are often less polished than mainstream toolkits, but more directly aligned with safety.

If you want both locomotion + safety

A common practical stack is:

• Simulator: MuJoCo, Isaac Gym, or Isaac Lab
• RL framework: Stable-Baselines3 / RLlib / Tianshou
• Safety layer:
  • action shielding
  • constraint penalties
  • constrained optimization (e.g., CPO-style)
  • model predictive safety filter
  • runtime verification / rule-based guards

Sample-efficient algorithms commonly used for locomotion

• SAC — very popular and sample-efficient
• TD3
• PPO — often stable, but usually less sample-efficient than off-policy methods
• CPO / constrained variants — if safety constraints are central

Practical recommendation

If your goal is robot locomotion with safety constraints, I’d look at:

1. Isaac Lab + SAC/PPO
2. Add constraint handling via:
   • penalties,
   • constrained RL,
   • or an external safety filter.
3. If you want a more “safe RL” academic setup, look for CPO implementations or constrained RL libraries and pair them with MuJoCo or Isaac.

If you want, I can also give you:

• a short list of specific libraries with links, or
• a recommended stack for quadruped/humanoid locomotion depending on your simulator and safety requirements.