acceptodds
Withdrawn from ICLR 2027

REA: A Routed Embodied Agent Runtime for Asynchronous Brain-Cerebellum Control

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

12 views

Abstract

Foundation models are becoming remarkably reliable at high-level reasoning, yet they remain too slow and costly to serve as fine-grained controllers. Conversely, fast policies like VLAs and whole-body controllers offer more efficient reflexes, but struggle to generalize. While pairing a deliberative brain (foundation models) with a reactive cerebellum (fast policies) seems natural for embodied systems, two obstacles remain: methods are built and scored on incompatible stacks, and their asynchronous coordination is difficult. For example, an agent that calls the VLA in the loop can stall the fast execution, while a VLA that calls the agent on errors tends to re-trigger fallbacks repeatedly or completely miss silent failures. To address this, we propose REA (Routed Embodied Agent), a unified runtime where a deliberative System 2, an embodied router System 1, and reactive System 0 skills cooperate under a single golden rule: no running skill waits on the brain. Instead of querying the brain on the critical path, REA employs a route memory that stores each skill's measured success ahead of time, a boundary-triggered judge gate that consults the brain only after a skill returns, and a shared-budget escalation ladder for bounded recovery. Evaluated on 3,000 held-out episodes across 60 LIBERO-Pro tasks, REA reaches 95.9% success and beats the best single-agent baseline in its own skill pool (94.7%) while reducing control steps by 25% and control time by 17%. Where skills fail silently, the judge gate adds 3.1 points over standard exception handling on unselected episodes, at 2.1 the time per success, and 15.3 points on a failure-heavy stress subset. REA also carries over to BEHAVIOR-1K, RoboTwin 2.0 and a real Unitree G1 humanoid, solving the most episodes of the compared methods on all three. Our findings offer an insight for embodied design: an embodied brain is best utilized when its decisions are made ahead of time and queried only at natural behavioral boundaries, rather than wedged into the fixed control path.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Agent, Efficient AI, Embodied AI

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_XVuMOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes REA, a three-level runtime comprising executable skills (System 0), a router (System 1), and an agent for skill authoring or completion judgment (System 2). Its main components are development-set route memory, a boundary-triggered judge, and a fixed escalation ladder. REA-M achieves 95.9% success on 3,000 LIBERO-Pro episodes versus 94.7% for the strongest single-author baseline, with additional evaluations on two simulators and a real humanoid.

Critical strengths
  1. The paper addresses the latency introduced by online agent calls during skill execution. REA provides a common runtime for skill execution and routing, allowing different configurations to be evaluated on the same tasks, seeds, and execution budgets.
  2. The LIBERO-Pro experiments cover 60 tasks with 50 held-out seeds per task. The paper reports paired statistical tests and ablations of route memory and the judge gate. It also examines the additional execution time associated with judge-based recovery.
Critical weaknesses
  1. Experimental Evaluation and Component design.

The contribution of the escalation ladder is unclear. Table 1 reports at least 95.7% success for route memory only (KK=0), compared with 95.9% for REA-M. In addition, route memory is fitted on development episodes of the same 60 tasks, while evaluation on unseen tasks is not provided. The results mainly support per-task skill selection, with limited evidence for the benefit of escalation or routing on new tasks.

The latency experiments do not isolate asynchronous scheduling. In Table 3, the blocking advisor makes 1.71 online model calls per episode, whereas REA-M makes none. This demonstrates the cost of online inference but does not establish an advantage over a non-blocking advisor with a matched inference budget.

The contribution of the judge gate is difficult to isolate from re-runs and replay variance. On SS-center, Fix and Evo are byte-identical, yet REA-V succeeds in 9/12 episodes compared with 4/12 for REA-M (Table 7). Table 11 uses REA-M as a proxy for the first-run outcome, making the attribution of rescued failures and replay variance approximate. A more direct analysis of the first-run outcomes is needed.

The experimental evaluation is insufficient. The RoboTwin 2.0 experiment does not evaluate the routing method used on LIBERO-Pro. Appendix J states that RoboTwin has no skill library. Instead, REA uses a reflex model as System 0, and the router hands single decisions to a deliberate model. Therefore, this experiment does not test the route memory and escalation ladder used in the main experiments.

  1. The authors employ a substantial amount of specialized terminology throughout the paper. Apart from terms that are standard and universally recognized in the field, most lack explicit definitions or supporting references, making the paper difficult to comprehend.

For example, the abstract uses "best single-agent baseline" for the 94.7% result, while Table 1 evaluates frozen programs from a single skill author. Figure 1 uses "first/other/first" to describe the escalation ladder without explaining this notation in the caption, although Section 3.2 specifies a different order when three candidates are available. The use of "asynchronous" also needs clarification, since Algorithm 1 waits for JUDGE to return before the next run.

Questions for authors
  1. Table 1 reports at least 95.7% success for route memory only (KK=0). What is the actual success rate when this configuration is evaluated independently? How does route memory perform on unseen tasks?
  2. Can the authors compare REA with an online advisor that does not block skill execution, using the same number of System 2 calls and inference budget?
  3. How much of the improvement from the judge gate comes from rescued failures, re-runs, and replay variance? Can the authors report the actual first-run outcomes instead of using REA-M as a proxy?
  4. Since RoboTwin 2.0 has no skill library, which parts of the routing method are evaluated in this experiment? Can the authors explain how these results support the conclusions drawn from LIBERO-Pro? completion labels and retry-matched baselines, and demonstrate the same routing mechanism across platforms?
Reviewer_C9qEOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes REA, an embodied runtime that separates a slow reasoning model, a skill router and fast execution skills. System 2 authors skills offline and, online, can either advise on dispatch or judge task completion from images. System 1 selects skills using per-task success rates and step counts measured on development episodes, and decides whether to stop or recover when a skill returns. System 0 executes frozen programs or other controllers. The three core components are a route memory fitted on development episodes, a visual judge gate triggered at skill boundaries, and an escalation ladder that shares one step and time budget.

The main experiments are on 60 LIBERO-Pro tasks, where the skills are programs written offline by coding agents. REA-M, which makes no online model call, reaches 95.9% success, about 1.2 points above the strongest single skill author (94.7%), with about 25% fewer control steps per success and 17% less control time per episode than that author. On the weaker ASPIRE skill library, adding the visual gate raises success from 71.9% (REA-M) to 74.9%, but the average time per successful episode is about 2.1 times that of REA-M. The authors also report a module ablation on a failure-enriched subset and portability experiments on BEHAVIOR-1K, RoboTwin 2.0 and a Unitree G1 humanoid.

Critical strengths
  1. A potentially useful empirical finding about where the brain should sit. On a program library, the paper finds that the online uses of the foundation model it tests add latency without improving success, while per-task selection on development episodes comes close to a hindsight oracle at no test-time model cost. This may be of interest to practitioners building code-as-policy systems.

  2. A broad comparison within a unified framework. The paper evaluates skills from several authors, routers with and without an online model, and different placements of the foundation model, all under the same episodes, horizon and wall budget, with 3,000 episodes per method on LIBERO-Pro.

Critical weaknesses
  1. The main result comes from per-task program selection, not from the asynchronous coordination the paper emphasizes. REA-M reaches 95.9% without any online model call, and route memory alone (K=0) already reaches at least 95.7% (Table 1), so the judge gate and escalation ladder contribute almost nothing here. REA's candidate pool combines programs from several authors, each requiring its own development effort, while each single-author baseline uses only one author's programs, so the 1.2-point gain over the best single author may not hold under a matched budget.

  2. The gain from the visual judge is not cleanly attributed, and the judge's overall accuracy is low. On the ASPIRE library, the judge adds about 3.1 points of success at about 2.1 times the time per success (Tables 12–13), but this comparison has no unconditional-retry control. The only such control, on the 72-episode matched subset (Figure 4), differs from the full router in its retry setup, so the paper cannot separate the judge's effect from that of extra executions. The judge's final-verdict accuracy on LIBERO-Pro (0.67) is also below that of always answering "done" (0.75).

Questions for authors
  1. Route memory alone (K=0) already reaches at least 95.7% (Table 1). Could you add a baseline in which the strongest single-author pipeline receives the same total development budget and selects its best program on development episodes? This would show whether REA's gain holds beyond additional development effort.

  2. Could you compare the visual judge with unconditional retry under the same first skill, skill sequence and retry limit, ideally on the 590-episode ASPIRE study where the +3.1 points are measured? The current control (matched subset, Fix-only, one retry) cannot separate the judge's effect from that of extra executions.

  3. How does the visual judge perform on the 179-program library used in the main experiment? Its final-verdict accuracy (0.67) is below that of always answering "done" (0.75), and its gains are reported only on weaker libraries and a failure-heavy subset. If it lowers success on a strong library, it should not be presented as a general mechanism.

  4. On RoboTwin 2.0 and the G1, were route memory, the judge gate and the escalation ladder actually enabled?

Reviewer_hnjqOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

REA is a runtime for dual-system embodied agents built around one rule: no running skill waits on the brain. A System 2 foundation/coding agent authors frozen skills offline and, online, only judges completed skills; a non-learning System 1 router dispatches skills using a route memory (per-task success rate and mean steps measured on development episodes), escalates through a fixed [first, other, first] ladder under a shared step/time budget, and optionally consults a VLM judge only at skill boundaries. On 3,000 held-out episodes over 60 LIBERO-Pro tasks, REA-M reaches 95.9% vs 94.7% for the best single skill author in its pool, with 25% fewer steps and 17% less control time; a judge-gated variant adds 3.1 points over exception handling on weaker skills. Portability is claimed on BEHAVIOR-1K, RoboTwin 2.0 and a real Unitree G1.

Critical strengths

(1) The main comparison is exceptionally well controlled — identical (task, seed) pairs, initial states, frozen skill bytes and budgets across arms, a single post-hoc success read, paired tests with Holm correction — and the paper discloses unfavourable facts itself (non-significant corrected subset results, replay variance, judge accuracy below always-done). The clean negative result that an agent at dispatch adds wait time but zero success over a lookup table is valuable. (2) Per-task development-set selection is cheap (zero online model calls), lands within 1.0 point of the hindsight per-task oracle, and the paper honestly localises the gain (route memory alone already achieves ≥95.7%).

Critical weaknesses

(1) Novelty is thin relative to the framing: route memory is per-task best-of-pool selection on a development set (classic algorithm portfolios; the paper itself cites outcome-based routing for coding agents), the judge is an existing VLM success detector repositioned to skill boundaries, and the ladder is fixed. No learned or similarity-based router is proposed or compared. (2) The method is fitted per task: held-out evaluation changes seeds, not tasks, and the paper states that on a new task REA falls back to the globally best author (94.7%), i.e. the headline gain vanishes by construction on unseen tasks. (3) The brain–cerebellum claim is not demonstrated where it matters: all main-benchmark skills are agent-written programs; no VLA or whole-body controller is ever run as System 0, and the portability evidence is 2 tasks/20 episodes (BEHAVIOR-1K) and one scripted task/10 trials (G1, p=0.35 vs the main baseline). (4) The judge-gate contrast is confounded by escalation budget (K=1 vs K=2) with no K-matched unconditional-retry arm, and the +3.1 points cost 2.1x wall time per success from a judge whose overall accuracy is below always answering "done" on LIBERO-Pro. Results also depend on non-public, version-unstable model endpoints (the judge model was replaced mid-campaign), so the headline numbers are not independently reproducible as submitted.

Questions for authors

(1) What is REA's success on tasks with no per-task route memory (leave-tasks-out), beyond falling back to the global best author? (2) With identical K, ladder and budget, how do judge gate, exception gate and unconditional try-next-skill compare on the unselected ASPIRE set? (3) How do results vary with development episodes per program (3/5/10/20) and across independent development splits? (4) Can you run REA with an actual VLA as System 0 anywhere, or will you scope the claims to program skills? (5) What is the end-to-end cost per success including skill authoring and route-memory development?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.