AI

HiRoC Post-Trains VLAs as Planner Plus Subgoal Executor

Robb Harlan 5 min read

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.

Flat VLA failing place-bowl versus HiRoC staged approach-lift-place subgoals
Flat global instruction vs stage-wise subgoals. Source: arXiv:2608.05999 HTML.

Three training stages

  1. Planner SFT — Qwen2.5-VL-3B (LoRA) on cleaned VLA-OS-style subgoal labels; then frozen.
  2. 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).
  3. 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).

HiRoC framework diagram for planner SFT, executor SFT, and RL tuning
Planner SFT, executor alignment SFT, then hierarchical GRPO. Source: arXiv:2608.05999 HTML.

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.

Real-world HiRoC experiment placing correction fluid in a box
Real deployment example from the paper. Source: arXiv:2608.05999 HTML.

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.

Sources