Robot Motion 101 (2/3) — From Classical Control to Learned Policies
From Motion Planning and Impedance Control to ACT, Diffusion, and RL
A practical bridge from classical planning and contact control to imitation, generative policies, and selective RL
First published: 2026-07-15 | Last updated: 2026-07-15
From Planner to Motor Loop
Connect policies, planners, IK, low-level control, and real-time interfaces through explicit ownership and failure boundaries.
From Demonstrations to Policies
Build a testable path from teleoperation data through BC, ACT, diffusion, and flow-based policies.
Deploy Without Removing the Classical Spine
Place RL and learned actions inside constraint projection, contact control, watchdog, fallback, and staged promotion gates.
Part I: Move Reliably First — Classical Planning and Control
Where Does a Robot Policy Run? — From Planner to Motor Loop
Map task intent, skill policies, motion generation, low-level control, real-time interfaces, and the safety supervisor into one execution hierarchy.
→ 02The Default for Structured Tasks — IK, Motion Planning, and Trajectory Generation
Explain why classical methods remain the default for structured tasks and separate the roles of FK, IK, collision avoidance, timing, and trajectory optimization.
→ 03Stabilize the Last Meter of Contact — Impedance, Admittance, and Force Control
Choose and validate position, velocity, torque, impedance, admittance, and force control through energy, latency, and contact-stability reasoning.
→Part II: Move from Demonstrations to Policies — Teleoperation, BC, ACT, and Generative Policies
Produce the Raw Material for Learning — Teleoperation and Dataset Engineering
Design episode schemas, independent splits, replay, and dataset QA that preserve synchronization, calibration provenance, failures, and interventions.
→ 05The Simplest Learned Policies — From Behavior Cloning to ACT
Use ACT as a practical reference for behavior cloning, covariate shift, compounding error, action chunking, temporal ensembling, and control latency.
→ 06Generate Multiple Plausible Actions — Diffusion Policy and Flow Matching
Compare diffusion and flow policies that generate conditioned action trajectories from multimodal demonstrations and execute them in a receding horizon.
→Part III: Add Learning Only Where Needed — RL, Hybrid Stacks, and Deployment
Where Does RL Fit? — Reward, Residual Policies, and Fine-Tuning
Define gates for simulation skill learning, residual policies, gain or trajectory adaptation, imitation fine-tuning, and offline versus online RL.
→ 08Do Not Insist on One Method — Classical + Learned Hybrid Architecture
Combine learned policies with planners, IK, constraint projection, contact control, and fallback while defining interfaces a future VLA must obey.
→ 09Build the Learning Stack One Gate at a Time — A Baseline-to-Policy Runbook
Take one tabletop assembly from a classical baseline through data, ACT, a generative policy, optional RL, and a bounded real trial.
→