Would GRPO's Group-Relative Baseline Work for a Robotics RL Task?

A colleague working on a warehouse-robotics picking policy (the same kind of task covered in ppo-and-modern-policy-optimization's opening scenario) proposes switching from PPO with GAE to GRPO, arguing "GRPO is strictly simpler since it drops the value network entirely — why would we keep paying for one?"

1. Would dropping the value network in favor of a group-relative baseline transfer cleanly to this robotics setting? Identify the specific structural difference between LLM-reasoning RL and this robotics task that determines the answer.
2. If the team tried GRPO here anyway, what would you expect to go wrong or become harder, specifically?
3. State the general rule for when a group-relative baseline is a good substitute for a learned value function, rather than treating "no value network" as an unconditional simplification.