Part IV · Ch. 34 — Embodied Robotics

draft

Hybrid · Embodied/control · Vision + RL + control + (LLM planner). AI that perceives and acts in the physical world, integrating several modalities. Card: ../02-types-of-ai/06-hybrid-neuro-symbolic.kmd.

🎨 Figure F-IV.34.0 — Perception → plan → physical action. Brief: an arm/robot with a camera perceiving objects, a plan in language ("pick up the cup"), and the control policy moving the joints; sensorimotor loop.

Embodied Robotics

1. Definition and short history

Unites perception (CNN/ViT, ch. 03-04), control policy (RL, ch. 26-28) and, recently, LLM planning. Lineage: classical control → deep RL for robotics (2016+) → Vision-Language-Action (VLA) models (2023+).

2. Foundations

  • Control theory / cybernetics — sensorimotor loop, stability.
  • Physics / mechanics — dynamics, kinematics, contact.
  • Probability / stochastic processes — sensory uncertainty, MDPs.
  • Geometry — 3D space, transforms, pose.

3. Algorithms and architectures

  • Perception — CNN/ViT, sensor fusion (camera, LiDAR, touch).
  • Policy — RL (PPO/SAC), imitation learning, behavior cloning.
  • Planning — LLM/symbols decompose tasks; world models.
  • VLA — single vision+language→action model.
  • Sim-to-real — training in simulation + transfer to the real world.

4. Inputs

  • Hardware: robot + sensors; GPU for training; embedded compute for inference.
  • Data: demonstrations, teleop, massive simulation; real data expensive.
  • Data structures: point clouds, maps, trajectories, states.
  • Systems: ROS, simulators (Isaac, MuJoCo), real-time control.

5. Specialized life cycle

Stage Specialization
0 Problem Physical task (manipulation, navigation), safety, environment
1 Data Demonstrations + simulation; domain randomization
2 EDA Scenario coverage, sim-to-real gaps
3 Modeling Perception + policy + planner; VLA?
4 Training RL/imitation in simulation; fine-tune on the real one
5 Evaluation Task success rate, robustness, safety
5.5 Acceptance Physical tests, safety limits, fail-safe
6 Production Real-time control; human supervision; physical guardrails
7 Monitoring Success, failures, wear, sensor anomalies
8 Retraining Field data, new objects/environments
9 Governance Physical safety, accountability, job displacement

6. Capabilities, modes, and modalities

Bodily-kinesthetic + spatial + executive: manipulation, locomotion, autonomous navigation, vehicles. The type that integrates the most modalities and acts in the physical world (unlike digital agents, ch. 30).

7. Limits, risks, and ethics

Sim-to-real gap; expensive real data; physical safety (risk to people); robustness to unseen environments; legal accountability; impact on jobs.

8. State of the art and examples

Learned manipulation, autonomous vehicles, humanoid robots, VLA models (RT-2-like); trend: robotics foundation models and greater generalization across tasks and bodies.

Hybrid/Neuro-symbolic Paradigm complete (ch. 30–34). Part IV index: INDEX.kmd.