Anyone training locomotion policies from scratch vs fine-tuning a foundation policy?

[zoeanderson]

Speaking from personal experience here, 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. 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.

[ramirez77]

Still learning the space, so correct me if wrong - 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.

[noah_pate]

I'll believe the stronger version of that claim when it's independently verified. 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.

[sarah.santos3]

@noah_pate Tangent, but worth mentioning: 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. 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.

[mia_lars]

@sarah.santos3 I'll believe the stronger version of that claim when it's independently verified. 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. 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. Totally unrelated but has anyone else noticed how fast component costs are dropping this year.

[wei_ross]

I'd frame this differently. 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.

[carol.robinson]

@wei_ross I'd take that specific number with a grain of salt, honestly. 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. Makes me wonder how this looks in another five years.

[rossi30]

@carol.robinson Side note that might be relevant: 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.

[deborah59]

@rossi30 Not to derail, but this reminds me of something adjacent: OpenVLA is a notable open-source VLA model - roughly 7 billion parameters, trained on hundreds of thousands of real-world robot demonstrations - and has been shown to outperform much larger closed models on some manipulation benchmarks, which says a lot about how much of VLA performance comes from data curation rather than raw scale. 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.

[ivan22]

@deborah59 Counterpoint: 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.