Why this guide
This guide turns a dense sim2real draft into a practical, implementation-first series for robotics policy deployment. The structure follows the path most teams actually execute:
- train a privileged teacher in simulation,
- distill and fine-tune a deployable student,
- calibrate simulation with system identification,
- and close the sim-to-real gap in hardware rollout.
Each chapter includes implementation scaffolds, TODO placeholders for IsaacLab code, and media placeholders for videos of each stage.
Scope. This series focuses on policy training and deployment for manipulation tasks using the teacher-student paradigm. We prioritize practical deployment decisions over exhaustive literature coverage.
End-to-end pipeline map
Phase A: Teacher training in simulation
-> Privileged observations
-> RL algorithm choice
-> Action space + resets + randomization
Phase B: Student distillation and finetuning
-> RGB + proprioception student
-> Distillation objective
-> Optional RL finetuning
Phase C: SysID and sim2real deployment
-> URDF and actuator calibration
-> Visual alignment
-> Runtime action and safety checks
How to use this guide. Read chapter-by-chapter if you are building a new pipeline, or jump directly to the chapter that matches your current bottleneck.
Chapter roadmap
Part 2: Teacher training
Build a robust privileged teacher policy with early algorithm selection, then refine with action-space design, reset strategy, exploration choices, and randomization.
- Link: Part 2 - Teacher Training
- Focus: RL algorithm selection, training stability, and policy robustness in sim.
Part 3: Student distillation and finetuning
Convert the teacher into an RGB student, choose model architecture, and optionally fine-tune for task-specific robustness.
- Link: Part 3 - Student Distillation and Finetuning
- Focus: DAgger/BC setup, visual encoders, and post-distillation improvements.
Part 4: SysID and sim2real deployment
Calibrate simulation to hardware and deploy safely with execution smoothing, validation, and runtime guardrails.
- Link: Part 4 - SysID and Sim2Real Deployment
- Focus: system identification, transfer reliability, and deployment playbook.
Recommended order. Start with Part 2 unless you already have a strong teacher. If a teacher is already available, jump to Part 3 and then Part 4.
Code and demo assets
Below is the shared scaffold we will expand across Parts 2-4.
# TODO(aditya): Shared experiment config scaffold for this series.
# Replace with your IsaacLab experiment registry and task definitions.
EXPERIMENTS = {
"teacher_training": "TODO: IsaacLab teacher config path",
"student_distillation": "TODO: IsaacLab student config path",
"sysid_and_deployment": "TODO: IsaacLab deployment config path",
}
Demo index placeholder: