BB.Policy

Learned policies for Beam Bots. bb_policy lets robots execute neural-network behaviours that map observations to actions — train a policy from demonstrations or simulation, export it to ONNX, and deploy it on real hardware with full safety-system integration.

A policy is a function π: observation → action. Given what the robot perceives (joint positions, velocities, camera frames, forces), the policy outputs what it should do (target positions, velocities, gripper commands). Inference runs on the BEAM, in the same runtime as control — so a crashed or slow policy can't take the robot down with it.

🚧 Early scaffold. The BB.Policy behaviour and project conventions are in place; implementations are landing in phases. See PROJECT_PLAN.md for the roadmap and the design decisions behind it.

def deps do
  [
    {:bb_policy, "~> 0.1"},
    # ONNX inference is optional — add ortex when you deploy a real model:
    {:ortex, "~> 0.1"}
  ]
end

{:ok, result} =
  BB.Policy.run(MyRobot, BB.Policy.ONNX, %{task: :pick_mug},
    policy_opts: [
      model: "priv/models/pick_mug.onnx",
      normalizer: "priv/models/pick_mug.json",
      observation: [positions: [:shoulder, :elbow, :wrist]],
      action: [{[:shoulder, :elbow, :wrist], :position}]
    ],
    rate_hz: 20,
    timeout: :timer.seconds(30)
  )

Generated docs live at https://hexdocs.pm/bb_policy. Architecture, decisions, and the phased roadmap are in PROJECT_PLAN.md; contributor conventions are in AGENTS.md.

Apache-2.0. See LICENSES/ and the SPDX headers on each file (REUSE-compliant).