DifrasAI: AI Research Company

03 · Backends

Configs & adapters

Swap sim, robots, vision, IK, policy, and language backends via YAML without rewriting the control loop.

Configs

File	Use
`configs/default.yaml`	Mock sim (CI / fast)
`configs/mujoco.yaml`	Physics + viewer
`configs/bitnet.yaml`	Optional 1.58-bit reasoner

Robot backends

Backend	Status
`mock_sim`	Default lightweight arm
`mujoco`	Tabletop physics + cameras
`ros2`	Joint trajectory (dry-run by default)

New arms: implement robots.base.RobotAdapter.

Optional backends

control.ik_backend: dls
Jacobian damped-least-squares IK
scene.backend: yolo
Pretrained detector (pip install ultralytics)
policy.backend: lerobot
Run pretrained HF VLA checkpoints (SmolVLA, ACT) through the loop (pip install "lerobot[smolvla]"; CPU works: ~27s per 50-action chunk)

BitNet (optional LLM)

Use Microsoft bitnet.cpp as Kinetic's Language Reasoner for 1.58-bit LLM inference on CPU (GPU kernels available upstream). Control / policy / vision stay separate, BitNet only emits intent JSON.

Property	How Kinetic uses it
1.58-bit weights	Ternary BitNet; official lossless kernels
CPU speedups	Upstream reports ~1.4–6×; large energy savings
Event-only	BitNet runs on new instruction / replan, never in the control loop
Compact context	Prompt = system rules + WorldState.summary_for_llm()
Fallback	Rule reasoner if BitNet is missing

Architecture placement

User instruction
      │
      ▼  (event only)
 BitNet b1.58  (run_inference.py)  →  GoalSpec JSON
      │
      ▼
 Task Planner → World Model → Policy → Controller → Robot

Never call BitNet every control tick.

Config sketch

language:
  backend: bitnet
  bitnet:
    repo_dir: F:/BitNet
    model_path: F:/BitNet/models/.../ggml-model-i2_s.gguf
    threads: 8
    n_predict: 128
    temperature: 0.1
    strict: false

One-shot setup

# Prefer VS2022 Developer PowerShell (clang + MSVC toolset)
cd F:\Kinetic
powershell -ExecutionPolicy Bypass -File scripts\setup_bitnet_light.ps1

This page covers optional backends for Kinetic (YOLO, BitNet, LeRobot, ROS2). Details match the in-repo BitNet setup notes.

Kinetic docs