Hierarchy & subgoals — Hi-VLA and CAC-VLA

Papers: What Matters in Orchestrating Robot Policies: A Systematic Study of Hierarchical VLA Agents (arXiv:2606.10267, Hu et al., June 2026 — “Hi-VLA” below) and CAC-VLA: Context-Gated Action Conditioning for Vision-Language-Action Models (arXiv:2607.04816, Xiong et al., July 2026). Banked from the 2026-08-07 lit slices; re-read at skim depth with number verification. Fed: #6 — the design constraints and the escalation map for the self-subgoal probe (pre-registered, opens at the first quiet local-GPU window).

Why this theme is live right now

Our aux-attribution result (#6: aux-off costs +0.462 panel MAE) and π0.5’s Implicit-HL finding agree that training on subgoal prediction shapes the action representation. The open increment is runtime hierarchy: decode a subgoal first, condition the actions on it. We own the seam — the [subgoal|…] conditioning slot — and the self-subgoal probe is pre-registered to measure exactly this at zero training cost. These two papers are the external map: one measures how much explicit hierarchy buys and where, the other shows what a production-grade version of “condition on your own predictions” looks like when the predictions are unreliable.

Hi-VLA — the systematic sweep

The setup: a frozen VLM planner (Gemini 2.5 variants) emits language subgoals; a frozen VLA controller (Gemini Robotics On-Device, 1–3B) executes them; the paper sweeps the orchestration choices — planner strength, thinking on/off, subgoal-refresh policy, what observation text the planner sees, and memory — over MuJoCo ALOHA suites (5 tasks × 3 categories × 200 trials) plus a real ALOHA.

The headline table is the cleanest quantification of where hierarchy pays that we know of:

On short-horizon tasks, naive hierarchy buys nothing (69.63 vs 69.57); on long-horizon it nearly triples flat performance. The gain is a long-horizon and indirect-instruction phenomenon.

Second finding we carried: refresh policy matters a lot. Success-detection termination is best (57.4% long-horizon), a fixed ~8 s window is close (52.4%), and letting the model predict its own horizon is worst (43.5%) — VLA stochasticity makes advance time prediction unreliable. Third: feeding the planner extra text (bounding boxes, contact info) beats raw images alone (38.8% → 52.4% long-horizon), with the stated failure mode that “VLMs tend to ignore image inputs as task becomes harder” — the hierarchy-side rendition of our #11 state-dominant-bias story.

Caveats: planner and controller are separate frozen models — the language interface is the only subgoal representation studied, and self-generated subgoals (one model planning for itself) are untested. Static environments, latency ignored.

CAC-VLA — conditioning on your own predictions, gated

CAC-VLA extends π0.5 with latent-action conditioning: a frozen pretrained action tokenizer encodes future action segments into compact latents; the VLM learns to regress those latents from learnable query tokens; the action expert consumes them through a learned channel-wise gate at every layer (a gated residual — the gate reads the current action states and the latent update and decides how much conditioning to let through). The asymmetry we banked it for: training conditions on ground-truth-encoded latents, inference on the VLM’s self-predicted ones, with no scheduled sampling — the gate is the stated mitigation for noisy self-predictions.

Numbers: LIBERO average 98.3% (vs π0.5’s 96.9% — a small +1.4 margin in a saturated regime); the meaningful gaps are LIBERO-Plus robustness (89.5% vs 85.7%) and real-robot (64% full success vs 16% for π0.5 on their pick-and-place; 25 trials/method). Ablations are thin: removing the gate costs only 0.4 points on saturated LIBERO; no gate-vs-cross-attention or λ sweep.

What transfers, and what it fed

Into the probe’s design (already frozen): Hi-VLA is why the probe’s per-step decomposition expects gains concentrated in late-horizon chunk MAE (their short-horizon null), and why any future rollout arm must pre-register its refresh rule (their worst-case was exactly the model choosing its own refresh). Since their planner/controller are separate models, our self-generated variant is a genuine increment, not a replication.

Into the escalation map (banked, not built): if the probe lands in the “oracle subgoals help but self-generated don’t” cell — the truth-vs-self asymmetry — CAC-VLA’s gated-conditioning pattern is the named escalation: train on truth, infer on self, let a learned gate modulate trust. Its thin ablations mean we’d treat it as a design sketch, not evidence of magnitude.

What doesn’t transfer: all their numbers are closed-loop success under a runtime planner loop; our panel probe conditions per-frame and sidesteps refresh policy entirely — which is precisely what makes it cheap, and also what it cannot measure (replan-timing effects are invisible offline).