HiRoC Post-Trains VLAs as Planner Plus Subgoal Executor
Most VLA post-training still treats the robot as a flat policy: one global instruction, one action stream. HiRoC (Hierarchical Robotic Control), posted August 6, 2026 as arXiv:2608.05999, splits that into a planner that emits subgoals and an executor that is RL-tuned to follow them.
Three training stages
- Planner SFT — Qwen2.5-VL-3B (LoRA) on cleaned VLA-OS-style subgoal labels; then frozen.
- Executor subgoal SFT — OpenVLA-OFT trained on reorganized (observation, subgoal, action-chunk) tuples so it does not cold-start on planner language (distribution-misalignment fix).
- Hierarchical GRPO — online RL with group-relative advantages mixing task-level and subgoal-level progress; only the executor updates.
Replanning runs periodically during rollouts (every 20 policy calls in their sim recipe).
Benchmarks
On LIBERO (Spatial / Object / Goal / Long), HiRoC posts strong suite averages versus OpenVLA, SmolVLA, VLA-RL, TGRPO, VAL-OS, and others. Headline numbers from their Table 1 include 95.6 / 98.0 / 96.0 / 98.0 on the four suites for HiRoC (they highlight 98% on Long). They also claim about a 10% average improvement over strong baselines in the abstract framing, plus better LIBERO-Plus zero-shot under seven perturbation types.
A sim-to-real case places correction fluid in a box on a real arm after training the executor in a matched simulator with JoyRA-0.1, no extra real fine-tune.
A Human’s Take
Long-horizon VLAs need intermediate semantic anchors more than another flat GRPO run. Freezing the planner is a practical trade — credit assignment across plan and act on sparse rewards is messy. I want to see how often replanning lies when the bowl is partially occluded; hierarchy only helps if the subgoals stay true.