REA: A Routed Embodied Agent Runtime for Asynchronous Brain-Cerebellum Control
Authors: Pengfei Zhou, Yixing Ma, Zhengfeiyang Zhang, Haoyuan Chen, Zhenglin Wan, Hexin Wang, Wangbo Zhao, Yang You
Organizations: National University of Singapore, HPC-AI Lab · InfRec, Cardinal AI Lab · University of California, Berkeley · The University of Tokyo · Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences · Hong Kong University of Science and Technology
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.
Figures & tables
Figure 2: REA across embodiments. Left: episodes in BEHAVIOR-1K, RoboTwin 2.0 and LIBERO-Pro. Middle: route memory and the judge gate recover from silent failures on a LIBERO-Pro test episode. An Evo–Fix–Evo sequence places the bowl on the stove after the judge rejects its first two attempts; no attempt raised an exception. Right: the same runtime makes a real Unitree G1 walk to and push a chair (8/10 trials).
Figure 3: Three ways to couple a slow brain with fast policies. (a) An agent that calls the fast policy puts its latency on every step. (b) A fast policy that calls the agent on errors is interrupted by persistent errors and never by silent failures. (c) REA dispatches from route memory, lets each policy run uninterrupted at native rate, and consults the System 2 judge only when there is a need.
Figure 4: Task success and control time on LIBERO-Pro. Left: Success rate versus control time per episode, with Wilson 95% confidence intervals for success rates. Higher success rates and shorter control times are better. Right: Control time per episode. Values in brackets show the change relative to the blocking-advisor configuration and all configurations are evaluated on the full LIBERO-Pro suite with 3,000 episodes. Each episode additionally incurs 2.6–2.7 s of reset time.
Method
Calls/ep
All 60 (%)
Aligned 37 (%)
Steps/succ ↓
single method
Codex + GPT-6
0
74.4
81.2
1470
ASPIRE Fix (59 / 36 tasks)
0
65.1
71.9
1566
Agents SDK + GPT-6 (best dev.)
0
94.7
93.7
1343
Random method per task
0
78.2
82.5
–
routed
Route memory only ( K=0 )
0
≥ 95.7
≥ 95.1
–
REA -M (memory + ladder)
0
95.9
95.3
1003
Table 1: Comparison on LIBERO-Pro full set (50 test seeds per task). Aligned (37 tasks) : REA uses exactly the same program files as the Codex, ASPIRE Fix, and Agents SDK baselines. For the remaining 23 tasks, its policy pool instead uses a harness-generated program in place of the Codex program, or an ASPIRE Evo program in place of the ASPIRE Fix program. The best method on development seeds 61–70 is Agents SDK (95.3%, vs. 75.0% Codex and 72.0% ASPIRE ). Both REA routers use the exception gate ( REA -M fails only 124 of 3,000 episodes).
Figure 5: Router ablation on the matched hard subset (6 tasks × 12 seeds = 72 episodes per method; identical initial states, GPU lanes and policy code). Without route memory the judge ladder is Fix, Evo, Fix; blind retry re-runs Fix once.
Figure 6: Where should System 2 sit? LIBERO-Pro seeds 1–10 (600 episodes per method). Left: success with Wilson 95% intervals; the two direct policies solve no episode (arrows at the axis edge). Right: seconds per episode spent in System 2 calls (agent, caption model or text judge), log scale. The direct policies act through 3 cm / 15 ∘ Cartesian IK steps and a gripper command, at most 80 decisions, from scene captions. Calls and paired counts in Table 5 .
Method
Clean
Randomized
Overall (%)
Agent and code baselines (historical runs)
Codex CLI
28/50
27/50
55
OpenAI Agents SDK
22/50
24/50
46
ASPIRE (RoboTwin port)
25/50
26/50
51
Show-Harness
5/50
1/50
6
REA (Code policies)
27/50
28/50
55
Table 2: Comparison with VLA and code policies on RoboTwin 2.0. Each row evaluates the same 50 clean and 50 randomized scenes. Success counts and overall rates are shown; best per column is bold. Historical baselines and the VLA extension were run in different windows; MEDex uses a separate host. Protocols and policy budgets are detailed in Section C.1 .
Figure 7: Results of portability experiments. Bars show success rates; whiskers indicate Wilson 95% confidence intervals. BEHAVIOR-1K: test instances 1–10 of five R1Pro pick-up tasks (50 episodes), 1,500 s per episode, gpt-6-astra for every method; Codex + GPT-6 and ASPIRE are frozen programs (the ASPIRE -style Fix loop), Agents SDK and Show-Harness are online agents. RoboTwin 2.0: the two test sets (200 scenes, 50 tasks, half clean and half randomized), at most 40 decisions per episode. Unitree G1 of real walk-and-push-chair trials and AlohaMini2 of real block-placement trials are both manually reviewed. Show-Harness was not run on either real robot due to the low success rates. Details in Sections C.2 , D.1 and D.3 .
Figure 8: Experimetal results of AlohaMini2 block placement. Success rates in two separate studies, with ten trials per method; labels give successful trials. (a) Four agent methods. (b) π0.5 alone and REA routing between π0.5 and Code. The two studies retain their own protocols and timing measures ( Sections D.3 and D.4 ).
Appendix figures & tables24 assets
Supplementary material from the paper’s appendix.
Appendix
Blocking advisor
REA -M
Suite
Succ
Control
Succ
Control
goal_swap
456
14.30
470
11.32
goal_task
435
14.19
442
9.94
object_swap
474
10.06
499
7.68
object_task
481
12.36
486
7.43
spatial_swap
491
14.75
491
11.02
Appendix
Table 3: Per-suite success (of 500) and mean control time (s) on the full set. REA -M is faster than the blocking advisor in every suite and at least as accurate in every suite.
Method
System 2 runs
Calls/ep
Succ. (%) ↑
b : c
Control (s) ↓
Δ
Wait (s) ↓
Agents SDK + GPT-6
offline
0
94.7
109:91
12.04
−10%
0
Blocking advisor
in the loop
1.71
94.1
–
13.33
–
2.62
REA -A
at dispatch
1.00
95.9
97:44
11.36
−15%
0.73
REA -M
offline
0
95.9
90:37
9.96
−25%
0
Appendix
Table 4: Speed on the LIBERO-Pro full set, ordered by where System 2 runs (3,000 episodes per method on one host; the routed methods share one 179-program library). Control: control time per episode, excluding reset; Δ : paired change in control time and b : c : episodes solved only by the method in the row : only by the blocking advisor; Wait: time per episode the control path is blocked on System 2.
System 2 placement
Configuration
Calls/ep
S2 time (s)
Success (%)
b : c
offline
REA -M (route memory)
0
0
91.0
–
dispatch + recovery
REA -A
0.99
0.64
91.5
6:3
+ completion
REA -A + agent judge
4.18
7.18
91.3
12:10
caption judge
text model
3.13
9.57
89.7
11:19
agent
3.23
7.94
88.8
9:22
direct policy
agent
52.9
121.8
0.0
0:546
Appendix
Table 5: Where should System 2 sit ( Figure 6 )? LIBERO-Pro seeds 1–10 (600 episodes per method, one host, legacy library, exception gate unless stated). b : c against REA -M; S2 time: seconds per episode spent in System 2 calls (agent, caption model or text judge). The per-step policies act through 3 cm / 15 ∘ Cartesian IK steps and a gripper command, at most 80 decisions, from scene captions.
Component
Dev. episodes
Input tokens
Dev. success of result
Codex + GPT-6 program generation
1,328
89.0M
75.0% (600 ep.)
Agents SDK + GPT-6 program generation
1,249
28.8M
95.3% (600 ep.)
ASPIRE Fix loop and Evo search
not recorded
not recorded
72.0% (590 ep.)
REA route memory (all 179 programs)
1,790
0
–
Appendix
Table 6: Development budgets on LIBERO-Pro (60 tasks). Program-generation episodes are the development evaluations the coding agents ran while generating their programs; route-memory episodes replay frozen programs on seeds 61–70.
Method
Succ
Rate (%)
Steps/succ
Wall/succ (s)
J calls/ep
Esc./ep
b : c
GPT-6 zero-shot, sample 1
116/600
19.3
4299
127
0
0
–
GPT-6 zero-shot, sample 2
91/600
15.2
5315
481
0
0
–
+ REA -V (both samples)
158/600
26.3
6023
317
2.45
1.45
50:8
ASPIRE Fix-loop policies
391/590
66.3
1557
40.9
0
0
–
ASPIRE Evo (11 tasks)
57/110
51.8
2529
78.4
0
0
–
+ REA -M (exception gate)
424/590
71.9
1467
57.9
0
0.01
49:16
Appendix
Table 7: LIBERO-Pro seeds 1–10: REA on policies generated by two System 2 methods ( Section 4.4 ). b : c = episodes solved only by REA : only by the better single library above it (paired exact McNemar, all p<10−4 ). Wall time includes judge latency; J = judge, Esc. = escalations. ASPIRE rows cover 590 episodes. Best per group in bold.
Figure 9: REA improves policy libraries of very different quality ( Table 7 ; bars are Wilson 95% intervals).
Figure 10: Success against control time per success. Left: LIBERO-Pro full set (one host; up and to the left is better). Right: ASPIRE policies on seeds 1–10 ( Table 7 ).
Suite
Task (abbreviated)
Seeds
Fix/12
Evo/12
libero_goal_swap
Open drawer, put bowl (GS-drawer)
31–39, 42, 44, 47
3
11
libero_goal_task
Open drawer, put cream cheese (GT-drawer)
31–42
4
5
libero_object_task
Pick up butter, place in basket (OT-butter)
31–39, 46, 47, 50
2
2
libero_spatial_swap
Bowl between plate and ramekin (SS-between)
8, 14, 31–40
4
12
libero_spatial_swap
Bowl from table center (SS-center)
31–41, 43
3
7
libero_spatial_swap
Bowl on ramekin (SS-ramekin)
31–40, 43, 44
5
3
Appendix
Table 8: Tasks of the matched hard subset. Five tasks have two distinct candidate policies ( ASPIRE Fix and Evo); on SS-center the Evo winner is byte-identical to Fix, so its two columns are two replays of the same program and differ only by replay variance.
Method
First policy
Ladder
Gate
Max escalations K
Single policy (Fix / Evo)
Fix / Evo
[Fix] / [Evo]
–
0
Blind retry
Fix
[Fix, Fix]
always
1
REA -V w/o route memory
Fix
[Fix, Evo, Fix]
judge
2
REA -M
route memory
route memory
exception
1
REA -V (full)
route memory
route memory
judge
2
Appendix
Table 9: Method configurations of the router ablation ( Table 10 ). All methods run the episode loop of Algorithm 1 ; blind retry uses the unconditional gate always .
Method
Memory
Gate
Succ/72
Succ. (%) ↑
95% CI
Steps/succ ↓
Single policy (Fix)
–
–
21
29.2
[19.9, 40.5]
4569
Single policy (Evo)
–
–
40
55.6
[44.1, 66.5]
2369
Blind retry
–
always
24
33.3
[23.5, 44.8]
6812
REA -V w/o route memory
–
judge
39
54.2
[42.7, 65.2]
4432
REA -M
✓
exception
37
51.4
[40.1, 62.6]
2657
REA -V (full)
✓
judge
48
66.7
[55.2, 76.5]
3510
Appendix
Table 10: Router ablation on the matched hard subset ( Figure 5 ; 6 tasks × 12 seeds = 72 episodes per method; Wilson intervals; identical initial states, GPU lanes and policy code).
Figure 11: Successes out of 12 per task; the policy that route memory dispatches first is shown under each task (* SS-center has a single distinct policy). REA -V is best or tied-best on every task.
Task
Method
Succ/12
Steps/ep
Steps/succ
Judge calls/ep
Escalation steps/ep
GS-drawer
Single policy (Fix)
3
2312
9246
0
0
Single policy (Evo)
11
2138
2332
0
0
REA -V w/o route memory
7
3581
6138
1.92
1267
Blind retry
6
3987
7974
0
1684
REA -V (full)
12
3846
3846
1.92
1682
REA -M (exception gate)
11
2138
2332
0
0
Appendix
Table 11: Full per-task statistics (steps include the 10 settle steps; escalation steps are the steps of every run after the first). On SS-center, Fix and Evo are the same program.
Task
Suite
Fix dev
Evo dev
First
Ladder
GS-drawer
goal_swap
1/10
10/10
Evo (higher rate)
Evo, Fix, Evo
GT-drawer
goal_task
1/10
1/10
Fix (tie; fewer steps)
Fix, Evo, Fix
OT-butter
object_task
0/10
3/10
Evo (higher rate)
Evo, Fix, Evo
SS-between
spatial_swap
1/10
10/10
Evo (higher rate)
Evo, Fix, Evo
SS-center
spatial_swap
6/20
= Fix
Fix (single policy)
Fix, Fix
SS-ramekin
spatial_swap
4/10
2/10
Fix (higher rate)
Fix, Evo, Fix
Appendix
Table 12: Route memory of the matched subset. Evo is dispatched first exactly where development shows it to be clearly stronger. On SS-center the Evo program is byte-identical to Fix, so both development configurations count as Fix evidence (6/20) and the ladder collapses to [Fix, Fix].
Figure 12: Judge calibration on development labels at threshold 0.6 ( done is the positive class). With the default prompt, showing the episode-initial frame lifts the judge above the majority-class baseline; without it, precision and recall collapse.
Prompt
Initial frame
Accuracy
Precision
Recall
Latency (s)
Cost/call (USD)
default
off
0.671
0.105
0.036
1.60
0.000247
default
on
0.792
0.587
0.661
1.60
0.000390
evidence
off
0.699
0.345
0.179
1.93
0.000353
evidence
on
0.694
0.447
0.750
2.01
0.000527
Appendix
Table 13: Development calibration sweep ( deepseek-v4-flash-vision-exp , threshold 0.6; the result is flat over 0.5–0.9). The frozen configuration is in bold.
Attribution (proxy = REA -M outcome)
Matched (72)
LIBERO-Pro (590)
Rescued: first run failed → escalated → success
7
28
Not rescued: first run failed → escalated → still failed
22
117
Survived: first run succeeded → judged not done → re-run → success
20
198
Harmed: first run succeeded → re-run → failed
1
19
Forfeited: first run failed → judged done → stopped
1
12
Kept: first run succeeded and judged done
16
207
Appendix
Table 14: Episode attribution of the full router. REA -M stands in for the first-run outcome, so the split of the paired wins into rescues and replay variance ( 12=7+5 ; 37=28+9 ) is approximate.
Method
trash
radio
teddy
drill
camera
Total
Prog. s/ep
Codex + GPT-6 (frozen program)
6
6
5
1
0
18/50
994
ASPIRE -style Fix (frozen program)
7
6
4
1
0
18/50
853
Show-Harness (online, scripted approach)
0
2
1
0
0
3/50
1492
OpenAI Agents SDK agent (online)
9
8
7
0
1
25/50
1049
REA (routed episode runtime)
9
7
5
1
5
27/50
1122
Appendix
Table 15: BEHAVIOR-1K, test instances 1–10 (successes of 10 per task; mean program time per episode). Every method uses gpt-6-astra . Paired against REA ( REA alone : the compared method alone): Agents SDK 7:5 ( p=0.77 ), Codex 13:4 ( p=0.049 ), ASPIRE -style Fix 12:3 ( p=0.035 ), Show-Harness 25:1 ( p<10−5 ). Best per column in bold.
Method (development instances 26–35)
Task
Succ
Steps/succ
Wall/succ (s)
Elapsed/succ (s)
Human-written reference program
trash
5/10
7795
1445
1565
Interactive agent (agent in the loop)
trash
4/10
–
–
6438
REA , frozen policy ( K=0 )
trash
7/10
6887
1611
1734
Human-written reference program
radio
2/10
27,947
5547
5848
Interactive agent (agent in the loop)
radio
4/10
–
–
7281
REA , frozen policy ( K=0 )
radio
4/10
9193
1945
2097
Appendix
Table 16: BEHAVIOR-1K development instances 26–35. Elapsed time includes reset and scene loading (for the interactive agent, per completed trial). The frozen policy is the best of three candidates on these instances, so its success is optimistic. Best per column in bold.
Condition
Episodes completed
Successes
Rate
Pro (6 suites, 600-step horizon)
1,987
1
0.0005
Appendix
Table 17: One-token VLM baseline (official test initial states; this policy uses its own 600-step horizon.
Method
Successful trials
Success (%)
p vs. REA
ASPIRE (code as policy)
2/10
20
0.023
Codex + deepseek-flash (Codex CLI)
4/10
40
0.17
OpenAI Agents SDK (tool agent)
5/10
50
0.35
REA
8/10
80
–
Appendix
Table 18: Unitree G1 walk-and-push-chair task, operator-judged success. Fisher exact tests against REA ; REA shown in Figure 13 .
Figure 13: REA on real robots. (a,b) Unitree G1 walks up to a chair and pushes it, shown from front and side external cameras; labels give time in each video. (c) AlohaMini2 block placement in a successful trial from the four-method study (trial 09): approach, lift, place and release, read left to right and then top to bottom. The first three images are external views; the last is the wrist view. Aloha labels give recorded observation numbers.
Stage
Original bridge
Fast harness
Service, single-shot
Service, pipeline
Capture
1.1–1.7 s
18 ms
14–36 ms
off the request path
Model
1.8–2.4 s
1.1–1.5 s
0.80–1.25 s
off the request path
Fused gate
2.3 s
4–10 ms
0.6–2.2 ms
0.6–2.2 ms
Motion accepted
1.5 s
7–45 ms
7–45 ms
7–45 ms (at execution)
Request latency
≈ 5 s
1.17–1.9 s
0.95 s
1.0–3.6 ms
Appendix
Table 19: Request latency on a Unitree G1 (medians over 5–8 trials). The two right columns are the same service; only pipelining differs. The pipeline total is request latency; the motion RPC (7–45 ms) and the live depth re-read (15–40 ms) follow at execution.
Method
Success
Rate (%)
Elapsed (s) all trials
Elapsed (s) successful
Actions per trial
Model calls per trial
Codex
5/10
50
1118.2
1699.2
13.6
32.1
Agents SDK
4/10
40
1079.5
1205.4
24.5
38.1
ASPIRE
8/10
80
1314.9
1222.2
30.1
35.7
REA
10/10
100
1618.0
1618.0
34.1
41.1
Appendix
Table 20: AlohaMini2 four-method study (40 valid trials). Elapsed time is the recorded elapsed_s . All means cover ten trials, including failures, except the successful-trial time column.
Method
Success
Rate (%)
Active (s) mean
Active (s) median
Elapsed (s) mean
Failures grasp / place
π0.5
3/10
30
16.57
16.34
29.36
2 / 5
REA ( π0.5 + Code)
8/10
80
128.89
46.29
132.70
2 / 0
Appendix
Table 21: AlohaMini2 VLA and Code study. All time statistics cover ten trials. Active time is activeMs /1,000; elapsed time is endedAt - startedAt and includes pauses.