robot-policy — an embodied policy in sim
The ETH Robot Learning course final project: a trained policy (imitation or reinforcement learning) in simulation, documented as a build story — the place where the robot stack and the LLM stack visibly converge (transformers-as-policies, RL, vision-language-action).
ExternalTrains in a physics sim with PyTorch on a GPU — lives in menfem-lab; results shown here as clips.
Stack
The problem
Robot learning and LLM post-training are the same RL body in two embodiments (tokens vs torques). Building one makes the other concrete.
The approach
Train a policy on a manipulation task in sim, starting from imitation and adding RL — reusing the PPO / reward intuitions from the Agentic AI rung.
Decisions & trade-offs
- —Start with imitation (behavioural cloning), then layer RL — the course arc.
- —Frame the write-up around the robot-stack ↔ LLM-stack convergence.
Where it stands
Target: a policy running in sim + a /notebook synthesis on why the two stacks are converging.
Build log
The running diary will appear here as /notebook entries once the build starts.