Steerability

The ability to change exactly how a model does something using prompts or conditioning, without retraining it.

Steerability originally comes from the large-language-model world, where it refers to whether a user can adjust a model's behavior at inference time through prompts or system instructions. Applied to robot foundation models, it means a policy not only knows what to do but can also be prompted to change how it does it. Physical Intelligence's π0.7, released in April 2026, treats this as a core selling point: during training, each data segment is paired with diverse context, including language describing the task and its sub-steps, metadata such as speed and quality, labels for the joint-space or end-effector control mode, and images of visual sub-goals. At inference time, these same kinds of conditioning can steer the model toward a different way of doing things, and it can even be taught a task it never practiced through step-by-step language guidance. This also means labeled failure data and suboptimal data become usable for training.

ExampleResearchers guided π0.7 through step-by-step language instructions, such as “open the air fryer with your left gripper” and “put the sweet potato in,” to operate a kitchen appliance it had never specifically been shown demonstrations for.

Also called

Steerable Policy

Related

π0.7 · Instruction Following · Language Corrections · Prompt / Prompt Engineering · Language-conditioned Policy · Goal-conditioned Policy

Sources

π0.7: a Steerable Generalist Robotic Foundation Model with Emergent Capabilities (arXiv 2604.15483)
π0.7: a Steerable Model with Emergent Capabilities (Physical Intelligence blog)

As of

2026-04