Anyone comparing PPO vs newer RL algorithms specifically for humanoid gait?

[charlesbianchi]

@zoeanderson Ran into exactly this myself. Whole-body control (WBC) formulates locomotion and manipulation as a single optimization problem across all joints simultaneously, respecting contact constraints and task priorities - it's more general than ZMP-only approaches but is computationally heavier and harder to tune.

[emma_whit]

@charlesbianchi Short answer: Model predictive control (MPC) is still very much alive in production humanoids, often working alongside or underneath learned policies - MPC handles short-horizon dynamically-consistent trajectory optimization while learned components handle perception, task-level decisions, or recovery behaviors that are hard to hand-model.

[rao91]

Same conclusion I've come to. Also worth noting: Isaac Lab (the successor to Isaac Gym) is widely used for large-scale parallel RL training thanks to GPU-accelerated physics, while MuJoCo is often used as a secondary 'sim-to-sim' validation step because its contact dynamics are generally considered more realistic than Isaac's, even though it trains slower at scale.

[erik_novi]

@rao91 Thanks for laying this out, genuinely useful. Physical Intelligence's pi0 pairs a smaller pretrained vision-language backbone with a separate flow-matching 'action expert' module, which is one way to get fast, high-frequency action output without needing the whole giant language model to run at control-loop speed. Reminds me a bit of the early drone hobbyist scene, honestly.

[mia_lars]

Here's what I know on this: Sim-to-real transfer still commonly breaks on contact dynamics - friction, restitution, and deformable/compliant surfaces are the hardest things to model accurately in simulation, so policies trained purely in sim often need real-world fine-tuning specifically around contact-rich tasks.

[tariqlarsen]

This lines up with my experience. ROS2 remains common in research and early-stage products for its tooling and ecosystem, but a number of production humanoid companies run custom, more tightly-optimized middleware for their real-time control loops, using ROS2-like tooling mainly for development, visualization, and non-real-time subsystems. Model predictive control (MPC) is still very much alive in production humanoids, often working alongside or underneath learned policies - MPC handles short-horizon dynamically-consistent trajectory optimization while learned components handle perception, task-level decisions, or recovery behaviors that are hard to hand-model.

[yuki71]

@tariqlarsen I don't think that's quite right, for what it's worth. A lot of what reads as 'full autonomy' in public demos is closer to a mix of scripted state machines, teleoperation for the hardest sub-tasks, and autonomous execution for the easier, well-rehearsed parts - transparency about this mix varies a lot between companies.

[deborahperez]

@yuki71 One nitpick - 'Zero-shot sim-to-real' rarely means literally zero real-world tuning in practice - it usually means the policy transfers well enough to be usable with only calibration and minor safety-limit adjustments, rather than needing a full additional training phase on hardware. Domain randomization - varying friction, mass, sensor noise, and even visual textures during training - is one of the more reliable tricks for improving sim-to-real transfer, but overdoing it can make training slower to converge and produce overly conservative policies.

[byang]

@deborahperez Slightly off-topic, but related: A lot of what reads as 'full autonomy' in public demos is closer to a mix of scripted state machines, teleoperation for the hardest sub-tasks, and autonomous execution for the easier, well-rehearsed parts - transparency about this mix varies a lot between companies.

[diego.moore6]

@byang Slight correction, though the overall point stands: Cross-embodiment training (training one policy across data from multiple different robot bodies) has shown some real transfer benefits for high-level behaviors, but low-level control (exact joint torques, timing) still tends to need embodiment-specific fine-tuning.