What's the current best practice for safe exploration during real-world RL?

[george92]

@olga_lind Follow-up question though - 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.

[shill]

This matches what I've seen too. Vision-Language-Action (VLA) models like RT-2, OpenVLA, and Physical Intelligence's pi0 unify a vision-language backbone with an action-output head, letting a robot map a camera image and a text instruction directly to motor commands instead of hand-coding separate perception and planning stages. 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.

[camila.jackson0]

One nitpick - 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. Balance-recovery controllers are usually evaluated with push-recovery tests (a known, repeatable lateral push) in demos, but real-world robustness also depends on recovering from unstructured events like uneven flooring, unexpected contact, or a dropped payload shifting the center of mass mid-stride - which is a much harder, less demo-friendly test.

[samuel.campbell8]

@camila.jackson0 This is exactly the kind of context I was looking for. 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. Diffusion policies model the distribution of possible actions and sample from it, which handles multimodal manipulation tasks (multiple valid ways to grasp something) more naturally than a single deterministic action output, at the cost of slower inference.

[aliu]

@samuel.campbell8 Worth being a little skeptical of the marketing angle here. 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. 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.

[forgecam45]

Slight correction, though the overall point stands: 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.

[shill]

This is a great summary, thanks. 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. '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.

[forgesve15]

This matches what I've seen too. '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.

[mohammed.rossi]

Same conclusion I've come to. Also worth noting: 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. 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.