PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Pengfei Zhou, Yixing Ma, Zhengfeiyang Zhang, Haoyuan Chen, Zhenglin Wan, Hexin Wang, Wangbo Zhao, Yang You

arXiv:2610.42053v1cs.ROcs.CV

TL;DR

Embodied agent and fast-policy methods lack a unified runtime and evaluation framework, while asynchronous coordination remains difficult. REA combines a deliberative brain, embodied router, and reactive policies in one runtime; across 60 LIBERO-Pro tasks, it achieves 95.9% success versus 94.7% for the strongest individual baseline, Agents SDK + GPT-6.

  • Problem

    Embodied agent and fast-policy methods lack a unified runtime and evaluation framework across simulators and real robots, while asynchronous coordination remains difficult.

    Source: §1 · p. 3

  • Method

    REA unifies a deliberative brain, embodied router, and reactive policies, using route memory, a boundary-triggered judge gate, and shared-budget escalation without blocking running policies.

    Source: §1 · pp. 3–4

  • Results

    Across 60 LIBERO-Pro tasks, REA achieves 95.9% success versus 94.7% for Agents SDK + GPT-6, the strongest individual baseline.

    Source: §1 · p. 3

  • Takeaways & Limitations

    The shared runtime supports learned VLA and programmatic code policies together, and its interfaces also carry over to simulators and a real Unitree G1.

    Source: §1 · p. 4, §6 · p. 14

    The re-run explanation is not a tested mechanism: its +15.3% subset gain is a stress test, while +3.1% on unselected seeds is the estimate to expect.

    Source: §4.5 · p. 11

Abstract

from arXiv · show

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 agentic 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 policies cooperate under a single golden rule - no running policy waits on the brain. Instead of querying the brain at each failure, REA employs a route memory that stores each policy's measured success ahead of time, a boundary-triggered judge gate that consults the brain on policy execution, and a shared-budget escalation ladder for bounded recovery. Evaluated on 3,000 test episodes across 60 LIBERO-Pro tasks, REA reaches 95.9% success and beats the best single-agent baseline in its own policy pool (94.7%) while reducing control steps by 25% and control time by 17%. In a separate 100-episode RoboTwin 2.0 study, REA achieves 84% success with VLA policies and 86% with VLA+code, compared with 80% for the strongest single VLA baseline MEDex. These results demonstrate learned and programmatic policies operating within the same runtime. REA also transfers to BEHAVIOR-1K and a real Unitree G1 humanoid, achieving the highest observed success among the compared methods.

1. Introduction

REA unifies deliberative agents and fast embodied policies in an asynchronous runtime where running policies never wait on the brain. Its routing, judging, and bounded-recovery mechanisms improve performance across LIBERO-Pro, RoboTwin 2.0, BEHAVIOR-1K, and real-robot trials.

Motivation

  • Foundation-model agents can plan embodied tasks but are costly, while VLAs scoring above 90% on LIBERO collapse on LIBERO-Pro’s perturbed test set; combining them faces runtime and coordination obstacles.An agent advisor that the router waits for leaves the robot idle for 2.6s per LIBERO-Pro episode (20% of its control time); an exception-triggered router escalated only 6 times in 590 episodes although 166 failed.Source: §1 · p. 3

REA runtime

  • REA combines System 2 agents, a System 1 router, and System 0 fast policies under a shared runtime and evaluation framework for code, VLA, and whole-body policies.Route memory uses per-task development success and cost; a boundary-triggered judge checks completion without blocking execution, and a shared step-and-time budget bounds escalation.Source: §1 · pp. 3–4

Results

  • 95.9% success across 3,000 LIBERO-Pro test episodes exceeded Agents SDK + GPT-6 (94.7%; p=0.004), with 17% less control time and 25% less than a blocking advisor.REA came within 1% of a hindsight per-task oracle.Source: §1 · p. 3
  • REA improves ASPIRE and zero-shot GPT-6 policies by +8.6% and +7.0%; judge gating adds +3.1% over exception gating (p=0.022), at 2.1× the time per success.Source: §1 · pp. 3–4
  • 84% with VLA policies and 86% with VLA+code in RoboTwin 2.0 exceeded MEDex’s 80% across 100 clean and randomized episodes.This study evaluated VLA and mixed VLA–code policies within the same runtime.Source: §1 · p. 4

Transfer

  • REA solved 27/50 BEHAVIOR-1K episodes versus 25/50 for Agents SDK (not significant), 97/200 RoboTwin scenes versus Codex CLI’s 75/200, and 8/10 Unitree G1 trials.Figure 2 depicts the runtime across simulators and a real Unitree G1 embodiment.Source: §1 · p. 4, Figure 2 · p. 4

2. Related Work

Prior work provides deliberative robot agents, fast embodied policies, dual-rate controllers, routers, and success detectors; REA combines these roles through policy-independent routing and escalation.

Embodied brains and fast policies

  • Deliberative agents and code-as-policy systems complement fast VLAs and whole-body controllers, which REA treats as off-loop System 2 reasoning and interchangeable System 0 policies.Examples include SayCan, PaLM-E, Code as Policies, and Voyager, alongside RT-2, OpenVLA, Octo, 𝜋0, and SONIC.Source: §2 · p. 4

Dual-system control

  • Unlike fixed-rate learned stacks such as GR00T N1 [13] and Helix 02, REA's actionless System 1 routes among policies while System 2 remains independent.Figure 3 contrasts stepwise brain calls and error-triggered fallbacks with uninterrupted policy execution and judge consultation only when needed.Source: §2 · p. 4, Figure 3 · p. 5

Routing, evaluation and completion judging

  • REA brings outcome-based routing to embodied policies, evaluates different method families on identical episodes, and uses existing detectors to gate escalation once per policy return, not define success.Prior routing includes FrugalGPT and Agent-as-a-Router [29]; Embodied Agent Interface and SIMPLER each evaluate one method family, while REFLECT turns detected failures into corrections.Source: §2 · p. 5

3. The REA Runtime

REA is a shared runtime for coordinating slow deliberative agents and fast policies. Its variants differ in how they select policies and trigger recovery.

3. The REA Runtime

  • REA-M uses route memory and exception-triggered recovery, REA-A uses a System 2 agent for dispatch and recovery, and REA-V adds visual judging at policy returns.REA-M makes no online System 2 calls; REA-V uses the judge to stop when done or escalate when not_done.Source: §3 · p. 5

3.1. Three-Level Contract

  • Policies execute closed loops at their native rate, while the runtime defines interfaces for policy runs, brain services, and backend scoring.Policies may yield and resume at segment boundaries; System 2 provides judging and advice, and each backend supplies a final score.Source: §3.1 · p. 6

3.2. The Embodied Router (System 1)

  • Route memory dispatches the candidate with the highest development success rate, breaking ties by lower mean steps, while REA-V judges each policy return.A shared-budget escalation ladder launches recovery runs from the current state; for RoboTwin, an observation-conditioned selector can choose among VLA policies or a mixed VLA-and-code pool.Source: §3.2 · p. 7

3.3. Aligned Evaluation Protocol

  • LIBERO-Pro comparisons align task-seed pairs, semantics, horizons, wall budgets, shared policy code, and final scoring; RoboTwin uses a fixed scene set and shared success criterion.Route memories and judge thresholds are set on development episodes before evaluation; RoboTwin policies retain their native action interfaces and budgets.Source: §3.3 · p. 7

4. Experiments

Experiments test REA’s routing success and efficiency, policy-generation methods, module contributions, VLA–code integration, and portability across simulators and real robots. The LIBERO-Pro full set covers 60 tasks and 3,000 episodes per method.

4.2. Routing Beats Each Policy-Generation Baseline

  • 95.9% success with route memory beats the best global method, Agents SDK, at 94.7%; REA also uses 1,003 steps per success versus 1,343 (−25%).On the 37 aligned tasks, REA-M reaches 95.3% versus 93.7%; on the other 23 tasks it is on par with the best method, 96.8 vs. 96.3%.Source: Table 1 · p. 8, §4.2 · p. 9

4.3. Route Memory Reduces Waiting for the Agent

  • REA-M cuts control time by 25% versus the blocking advisor and by 17% versus Agents SDK + GPT-6, while improving success by 1.8% versus the advisor.The blocking advisor leaves the control path idle 2.62 s per episode and delays the first action by 1.5 s; REA-A matches REA-M’s success but is 1.39 s slower per paired episode.Source: §4.3 · pp. 9–10

4.4. REA Improves Policies from Different Methods

  • On ASPIRE Fix-loop policies, REA-M raises success from 66.3% to 71.9%, and REA-V reaches 74.9%; on zero-shot GPT-6 policies, REA-V raises success from 19.3% to 26.3%.On Fix-loop policies, REA-M escalated only six times despite 166 failures, while the judge gate added 3.1% success at 2.1× wall time per success.Source: §4.4 · p. 10

4.5. Where Do the Gains Come From?

  • On the 72-episode Fix-failure subset, the full router solves 48/72, removing the judge gate costs 15.3%, and blind retry solves 24 versus 21 episodes.On the strong library, dispatching with an agent changes success by +0.5%, while completion judges cost 1.3–2.2%; the proposed explanation for judge gains is precise “done” decisions and selective reruns, but it is not a tested mechanism.Source: §4.5 · pp. 10–11

4.6. VLA and Code Policies Share System 0

  • REA reaches 84% success with VLA policies and 86% with VLA plus code, versus 80% for the MEDex single-VLA baseline.The mixed pool adds two randomized-scene successes, raising them from 40/50 to 42/50, with no lost VLA-only successes; a learned selector dispatches policies for each episode.Source: §4.6 · p. 11, Table 2 · p. 11

4.7. Extensional Experiments on BEHAVIOR and RoboTwin-2

  • REA solves 27/50 BEHAVIOR-1K episodes, the most among compared methods, and solves 97 versus 75 for Codex CLI in the side-by-side 200-scene RoboTwin transfer study.REA’s BEHAVIOR-1K margin over Agents SDK is not significant; the later-night RoboTwin results vary by 36 scenes between runs, so only the side-by-side pair compares methods without a window effect.Source: Figure 7 · p. 12, §4.7 · p. 12

4.8. Real Robot Experiments on Unitree-G1

  • REA succeeds in 8/10 Unitree G1 walk-and-push-chair trials, versus 5 for Agents SDK, 4 for Codex, and 2 for ASPIRE.All methods use the same model and System 0 policies, and success is judged on site.Source: Figure 7 · p. 12, §4.8 · p. 12

4.9. Real-World Block Placement on AlohaMini2

  • REA completes 10/10 AlohaMini2 block-placement trials, versus 8/10 for ASPIRE, 5/10 for Codex, and 4/10 for Agents SDK; in a separate controller study, it succeeds in 8/10 versus 3/10 for π0.5 alone.In the controller study, placement failures fall from five to zero while both groups have two grasp failures; REA’s mean recorded time is longer in both studies.Source: Figure 8 · p. 13, §4.9 · p. 13

5. Discussion

The discussion recommends routing with the brain before or between policy executions, not during a running policy. It also describes a shared policy interface that supports both route-memory lookup and learned selection.

Discussion

  • Route-memory dispatch outperforms each individual policy-generation method and matches asking an agent at every dispatch, with no online cost.The comparison is reported in Table 1 and Figure 6.Source: §5 · p. 13
  • 20–47% differences in success rates across methods show that the brain sets the ceiling while the router determines how much of it is reached.Choose the gate per library using development episodes: the free exception gate suits rare policy failures, while the judge gate pays off when failures are silent.Source: §5 · p. 13
  • A brain call inside the loop adds waiting, whereas a judge called when a policy returns can catch silent failures at extra time.The passage cites Figure 4 for waiting and Table 7 and Figure 5 for the judge-gate tradeoff.Source: §5 · p. 13
  • In the RoboTwin 2.0 extension, a learned selector uses the shared Run interface to choose one of seven VLAs or a Python policy from the initial observation and holds that choice fixed throughout the episode.The study evaluates the shared policy interface and initial selection; recovery-gate effects are evaluated separately on LIBERO-Pro.Source: §5 · p. 13

6. Conclusion … A.2. Same-Pool Baselines and Development Cost

REA’s router and shared runtime enable fast policies to operate without waiting on the brain, while supporting learned and programmatic policies across benchmarks and a real humanoid.

6. Conclusion

  • 84% success with VLA policies and 86% with VLA+code exceeded MEDex’s 80% on RoboTwin 2.0; REA also solved the most episodes among compared methods across three transfer settings.The transfer settings were BEHAVIOR-1K, RoboTwin 2.0, and a real Unitree G1.Source: §6 · p. 14
  • REA beats every method contributing policies to its LIBERO-Pro pool, and its gains come from routing rather than a larger policy pool.Route memory dispatches as well as an online agent at no model cost, while a boundary-triggered judge helps detect silent policy failures.Source: §6 · p. 14

A.1. Speed Details

  • 3.37 s less control time than the blocking advisor comprises 2.62 s less System 2 waiting and 0.75 s from 83 fewer control steps per episode.All timed methods ran on one host and operated at the same 9.5–10.4 ms per control step outside System 2 waits.Source: §A.1 · p. 19
  • Control time covers reset-to-episode-end, includes policy runs and waited-for System 2 calls, and is charged against a 900 s budget; wait measures time blocked on System 2.Intervals resample the 60 tasks over 4,000 bootstrap draws; 12 blocking-advisor episodes with no completed advice remain included.Source: §A.1 · p. 19

A.2. Same-Pool Baselines and Development Cost

  • At least 2,871 of 3,000 episodes succeed with K=0 route memory; REA-M’s 16 escalated episodes count as failures, though five succeeded after escalation.The first run raised an exception in each of those 16 episodes.Source: §A.2 · p. 19
  • The same-pool comparison uses the expected per-task success of uniformly choosing among three methods and an oracle selecting each task’s best test-episode method.On 37 tasks, the pool contains the exact programs used by the single-method rows; on the other 23, it includes generated Codex or ASPIRE Evo programs.Source: §A.2 · p. 19
  • Table 5 reports System 2 placement comparisons on LIBERO-Pro seeds 1–10, with 600 episodes per method on one host and System 2 time measured per episode.It also specifies the per-step policy actions and an 80-decision limit.Source: §A.1 · p. 19, Table 5 · p. 20
  • Table 6 compares development budgets on 60 LIBERO-Pro tasks, separating coding-agent program-generation evaluations from route-memory replays of frozen programs on seeds 61–70.The accompanying text states that Table 1 compares REA with each source method alone on the same 3,000 episodes.Source: §A.2 · pp. 19–20, Table 6 · p. 20

A.3. LIBERO-Pro Cross-Method Study … B.1. Episode Loop

REA-M ties for the highest success and is the cheapest timed method on the full set; on ASPIRE policies, route memory and the judge gate trade higher success for higher cost. Its episode loop uses fixed policy routes and shared budgets so escalation and judging do not extend an episode.

A.3. LIBERO-Pro Cross-Method Study

  • For zero-shot policies, REA-V routes over both samples; for ASPIRE, it routes over Fix programs and the Evo winner on 11 tasks, with one task excluded because its program calls sandbox-refused getattr.Table 7 covers LIBERO-Pro seeds 1–10; its ASPIRE rows cover 590 episodes.Source: §A.3 · p. 20, Table 7 · p. 21
  • Escalation averages 0.01 per episode under the exception gate and about once per episode under the judge gate, which accounts for REA-V’s extra wall time.Wall time includes judge latency.Source: §A.3 · p. 20

B.1. Episode Loop

  • Route memory fixes the policy ladder before reset, and each run receives only the remaining step and wall budgets, with judge latency charged to the same wall budget.Thus, escalation does not extend the episode.Source: §B.1 · p. 20

A.4. Success, Cost and Latency

  • On the full set, REA-M ties for highest success and is the cheapest timed method; on ASPIRE policies, route memory adds success at lower cost, while the judge gate adds more success at about twice the time per success.Figure 10 compares success against control time per success.Source: §A.4 · p. 20, Figure 10 · p. 21

B.1. Episode Loop

  • The success predicate is read once after the agent stops; before then, the loop does not read success, rewards, or object poses, and the judge’s verdict after the last run is only logged.The final score is recorded after stopping.Source: §B.1 · p. 22
  • The loop runs the selected policy with remaining budgets, then the judge gate evaluates the initial and current renders with the task label; judge completion or an exception-gate failure without timeout ends the run.The episode loop supports judge, exception, and always gate settings.Source: §B.1 · p. 22
  • In the matched hard subset, five tasks have distinct ASPIRE Fix and Evo candidates; on SS-center, Evo and Fix are byte-identical programs, so the columns differ only by replay variance.The two SS-center columns are replays of the same program.Source: Table 8 · p. 22

B.2. Matched Hard Subset: Tasks, Methods and Totals

The matched subset tests router ablations on six LIBERO-Pro tasks where ASPIRE’s Fix policy often fails, using shared initial states, GPU lanes, and policy code to compare methods that vary in their first policy, ladder, and gate.

B.2. Matched Hard Subset: Tasks, Methods and Totals

  • The subset includes six LIBERO-Pro tasks with twelve test seeds each, eight of which are Fix failures.The tasks were selected because ASPIRE’s Fix policy often fails on them.Source: §B.2 · p. 22
  • Router ablation methods run Algorithm 1 on the same initial states, GPU lanes, and policy code, differing only in first policy, ladder, and gate.Table 10 reports the totals, but the supplied passage does not provide their values.Source: §B.2 · p. 22

B.3. Per-Task Results on the Matched Hard Subset … C.1. RoboTwin VLA and Code Policy Integration

On the matched hard subset, the full router is best or tied-best on every task; on RoboTwin, mixed selection adds successes by choosing code on complementary scenes without mid-episode switching.

B.3. Per-Task Results on the Matched Hard Subset

  • REA-V is best or tied-best on every matched-hard-subset task; escalation improves tasks where the first policy is not strongest, at the cost of more steps.Where route memory dispatches the stronger policy first, the router matches or exceeds it.Source: §B.3 · p. 22, Figure 11 · p. 23

B.4. Route Memory

  • Route memory ranks policies by development success, breaks ties by fewer steps per success, and stores a [first, other, first] ladder; on SS-center, identical Fix and Evo programs collapse it to [Fix, Fix].It is fitted on development seeds 51–60, never test seeds; Evo is dispatched first where development shows it is clearly stronger.Source: §B.4 · p. 22, Table 12 · p. 24

B.5. Visual-Judge Calibration

  • Visual-judge calibration on 216 development labels found that showing the episode-initial frame lifts the default-prompt judge above the majority-class baseline; performance is flat across thresholds 0.5–0.9.The successor model scored 0.782 on the same labels, versus 0.792 for the frozen snapshot.Source: Figure 12 · p. 25, §B.5 · p. 25, Table 13 · p. 25

B.6. Mechanism of the Judge Gate

  • On test episodes, the final judge verdict had done-precision 0.83 on the matched subset and 0.905 on LIBERO-Pro.The matched-subset confusion matrix was TP 35 / FP 7 / FN 13 / TN 17; LIBERO-Pro was TP 276 / FP 29 / FN 166 / TN 119. Counts reflect only the verdict after the last run.Source: §B.6 · p. 25

C. Other Simulators and Online Baselines

  • Episode attribution uses REA-M as a proxy for first-run outcomes, so the paired-win splits are approximate: 12 = 7 + 5 on the matched subset and 37 = 28 + 9 on LIBERO-Pro.The splits distinguish rescues from replay variance.Source: Table 14 · p. 26

C.1. RoboTwin VLA and Code Policy Integration

  • The RoboTwin study reevaluated 100 fixed scenes using the official success criterion, and the selector chose among seven VLA policies plus a code policy at the initial observation.The code policy used four RGB views and robot state to generate one Cartesian move_gripper target per decision, with at most 40 decisions per episode. Its success head was trained on 247 scenes, selected using 40 development scenes, while the encoder, LoRA, and VLA heads remained frozen.Source: §C.1 · p. 26
  • The mixed selector chose code on nine randomized scenes and succeeded on eight; relative to VLA-only routing, it added successes on two tasks with none lost.On the other 91 scenes, its selected VLA and outcome matched the VLA-only run; this study used neither mid-episode switching nor a visual judge gate.Source: §C.1 · p. 27

C.2. BEHAVIOR-1K and RoboTwin 2.0 · C.3. Show-Harness Baseline

The sections report cross-platform evaluation protocols, paired RoboTwin comparisons favoring REA, and a zero-shot Show-Harness baseline with low success on LIBERO-Pro and RoboTwin 2.0. BEHAVIOR-1K preserves the LIBERO scoring contract while changing the backend.

C.2. BEHAVIOR-1K and RoboTwin 2.0

  • On BEHAVIOR-1K, the three-level contract, episode runner, and final ASPIRE task_completed() score remained unchanged from LIBERO; only the backend differed.The tasks were trash, radio, teddy, drill, and camera, using R1Pro 2025-challenge instances in Isaac Sim 4.5 and OmniGibson 3.7.2.Source: §C.2 · p. 27
  • BEHAVIOR-1K test instances 1–10 used one RTX PRO 5000 host, a 1,500-second per-episode program budget, and the same tools and domain text for online agents.REA used its RoboTwin-fixed routed policy for at most 40 decisions; the Agents SDK agent used at most 60 turns, and both used gpt-6-astra at medium reasoning.Source: §C.2 · p. 27
  • REA alone beat the compared method in paired RoboTwin runs: 46:16 vs Agents SDK, 40:12 vs ASPIRE, 66:8 vs Codex, and 121:0 vs Show-Harness (p<0.001 each).The side-by-side run against Codex was 42:20 (p=0.007).Source: §C.2 · p. 27

C.3. Show-Harness Baseline

  • Show-Harness [43] is a zero-shot VLM harness that plans per-arm subgoals and requests one discrete move per arm at every step using front and wrist views.Its unchanged code was run with gpt-6-astra and a lab-supplied embodiment adapter; development episodes set the adapter but were excluded from results.Source: §C.3 · p. 28
  • On LIBERO-Pro, Show-Harness achieved 78 successes (2.6%) across 3,000 episodes, with 2,806 episodes ending at the 900-second budget.Provider-failure episodes were rerun; the single-arm setup used high effort, at most 200 controller decisions, and 2 cm fine and 4 cm coarse moves.Source: §C.3 · p. 28
  • On RoboTwin 2.0, Show-Harness achieved 12 successes in 200 episodes in dual-arm zero-shot mode with its own 100-step budget.The harness used medium effort.Source: §C.3 · p. 28
  • For BEHAVIOR-1K, Show-Harness assumes a fixed-base arm at a table, so R1Pro approach and navigation were scripted before the harness started; grasp and movement used harness-planned grasping and inverse-kinematics steps.The scripted approach used the same search and navigation policies as other methods, without model calls, and remained inside the budget.Source: §C.3 · p. 28

C.4. An Agent as Its Own Cerebellum: the One-Token VLM Baseline … D.2. Unitree G1: Latency Check

The one-token VLM failed as a continuous LIBERO controller because it could not ground world axes in camera views, so REA uses it for policy-boundary judgments. The Unitree sections describe a chair-push comparison and a separate latency-only measurement.

C.4. An Agent as Its Own Cerebellum: the One-Token VLM Baseline

  • Two lateral tokens accounted for 89% of outputs, descend was almost never selected, and the end effector stayed more than 25 cm from the target.The sole success was a lateral sweep that flipped a stove knob.Source: §C.4 · pp. 28–29
  • A scripted controller completed pick-and-place tasks on 7 of 7 seeds with privileged object poses, indicating the failure was grounding rather than actuation.The scripted controller used the same tokens, step size, and horizon.Source: §C.4 · p. 29
  • The richer interface still solved 0 of 600 LIBERO-Pro episodes, so continuous control is left to System 0 policies.The richer interface used 3 cm / 15° Cartesian IK steps, a gripper command, scene captions, and up to 80 decisions.Source: §C.4 · p. 29

D.1. Unitree G1: Walk to and Push a Chair

  • The real-robot study tested a gantry-secured Unitree G1 on the instruction to walk forward, push an office chair, retract its arms, and stand still.Chair contact and displacement were not measured automatically; an operator judged and recorded each trial’s score on site.Source: §D.1 · p. 29
  • ASPIRE, Codex + deepseek-flash, the Agents SDK agent, and REA were compared using the same robot, model, and four System 0 policies.ASPIRE made one model request, Codex + deepseek-flash ran as a separate campaign, the Agents SDK agent made 2–3 requests, and REA made 4 policy-boundary decisions.Source: §D.1 · p. 29

D.2. Unitree G1: Latency Check

  • The Unitree latency service pipelines one-token vision decisions ahead of requests, with a live depth gate rechecking decisions at execution; this measures latency, not task success.Decisions were 0.04–0.79 s old; the gate’s 15–40 ms recheck was excluded from request latency, and reported medians covered 5–8 trials on the Jetson with live sensors.Source: §D.2 · p. 29

D.3. AlohaMini2: Four-Method Comparison · D.4. AlohaMini2: VLA and Code Policies

In the AlohaMini2 block-placement studies, REA completed more trials than the compared alternatives in both a four-method comparison and a separate π0.5-versus-Code routing study, while its runtime was longer where reported.

D.3. AlohaMini2: Four-Method Comparison

  • The four-method study tested moving an object into a box, requiring release and gripper retraction while the object remained stable inside.Codex, Agents SDK, ASPIRE, and REA each had ten valid manually reviewed trials; matching trial numbers used paired layouts and resets, with at most three grasp attempts and no total time limit.Source: §D.3 · p. 29
  • One successful four-method-study trial is illustrated as approach, lift, place, and release in Figure 13(c).The figure shows trial 09, with external views followed by a wrist view.Source: Figure 13 · p. 30
  • REA completed 10/10 AlohaMini2 trials, versus ASPIRE’s 8/10, Codex’s 5/10, and Agents SDK’s 4/10.The higher observed success accompanied longer mean runtime, more actions, and more model calls; early termination of failed trials affected runtime averages.Source: Table 20 · p. 31, §D.3 · p. 31

D.4. AlohaMini2: VLA and Code Policies

  • REA completed 8/10 trials versus 3/10 for π0.5 alone, with placement failures falling from five to zero and two grasp failures in each group.Both mean active time and mean elapsed time were longer for REA.Source: Table 21 · p. 31, §D.4 · p. 31
  • The separate study compared π0.5 alone against REA routing between the same step-5,000 checkpoint and a Code controller.Both used 15-target action blocks at 30 Hz; π0.5 predicted 50 steps per inference, and all ten REA trials used asynchronous routing. The groups came from different dates and task setups, with two invalid π0.5 records excluded.Source: §D.4 · p. 31
  • REA dispatched 449 π0.5 action blocks and 116 Code blocks; Code executed in five trials, while all ten counted as routed-configuration trials.The π0.5-only group dispatched 272 action blocks.Source: §D.4 · p. 31

Related papers

Loading 2610.42053v1…