ROBO·COP

Co-Evolving Robot Orchestrators and Policies through Deployment

Xilun Zhang1 Maggie Wang1 Erik Bauer1 Hong-Xing Yu1
Huang Huang1,† Jiajun Wu1,† Marco Pavone1,2,†

1 Stanford University 2 NVIDIA Research

† Equal advising

arXivCode (coming soon)

How can a robot deploy zero-shot in a new environment
and learn while performing its task?

Abstract

Vision-language-action (VLA) policies trained on large datasets are capable within their training domains, yet they still fail to generalize to the variety of situations a robot meets in real-world deployment. Agentic robot systems complement the policy with a vision-language model (VLM) orchestrator that learns when to call the policy, how to instruct it, and when to use scripted skills instead. However, because the harness is built around a frozen policy that has limited language steerability, the orchestrator can avoid the policy's failures but never overcome them. The policy becomes the bottleneck of the whole system. Fine-tuning the policy can remove this bottleneck, but updating it alone decouples it from an orchestrator tuned to its old behavior. We propose Robo-COP, in which the orchestrator and policy co-evolve during deployment. Robo-COP curates skill demonstrations from its own executions, fine-tunes the policy when this data can address recurring failures, and adopts each new policy only after it improves the skills it was trained for. Across ten simulated RoboLab tasks, Robo-COP raises mean held-out success from 64.8% to 73.8% over the same harness with a frozen policy, while fine-tuning on a fixed schedule without verification reaches only 65.8%. On three real-world tasks, Robo-COP raises held-out success from 38.3% to 50.0%. Robo-COP turns deployment into a self-improving flywheel in which robots learn by doing, with each improvement in execution producing better data for the next round of learning. Videos and code are available at https://xilunzhangrobo.github.io/robo-cop/.

Method

Robo-COP updates both the VLA policy and the adaptive harness that orchestrates its use.

The harness logs policy and scripted actions as one stream, curates it into skill demonstrations, and verifies each new policy before adoption, revising the memory tied to the old one.

Method — paper Section 3 ↗
Robo-COP inner loop updates orchestration memory; the outer loop curates data, trains and verifies candidate policies, and revises memory after policy selection.
The inner loop adapts the harness through episode reflection. The outer loop uses collected experience to update the policy, then revises memory so the harness can reassess the policy’s capabilities.

Skill demonstration curation

Recorded text and video judgments for three execution segments.

Execution calls

        
1. Text judge 
2. Video judge 
Play the segment to reveal the recorded video judgment.
3. Final decision
Awaiting video review

        
Original judge outputs; example verdicts are not an independent assessment of judge accuracy. Source records.

      

    


    
Simulation results
Robo-COP reaches 73.8% mean held-out success, versus 64.8% for the frozen-policy baseline and 65.8% for the fixed schedule.
Held-out success over ten simulation tasks and the task mean for five methods. Robo-COP has the highest plotted mean.
50 shared held-out initializations per task. Equally weighted ten-task means; whiskers are 95% beta-posterior credible intervals over the fixed task set. Paper Section 4.


    

      
Deployment and policy update cycle

      
Select a stage to view the corresponding execution record.

      

        
        
        
        
        
        
        
      

      

      
Step 1 of 6. Select any component to inspect it.

      
    


    
Simulation deployment
Videos, generated programs, and controller decisions from a 100-trial deployment run.


    

      

        
Simulation / 100 trials
Bananas out of bin
Training starts after trial 45; the adopted policy serves from trial 71.

        
BEFORE TRAINING · TRIALS 1–4551.1%
23 / 45 completed
CANDIDATE EVALUATION · 46–7068.0%
17 / 25 completed
AFTER ADOPTION · 71–10086.7%
26 / 30 completed

        

          

            
Outside VLA callCode 1
Early deployment
TRIAL 08 / 100
Scripted execution
The generated program samples a grasp and uses scripted motions to move the bananas.
Task completed

            
          

          
100 deployment trials
Rolling success rate with every trial shown below. Choose a 10- or 20-trial window.
 success failure candidate evaluation

            

          

          

        

        
Phase success rates from one deployment run.

      

    


    

      
Real world / 60 seconds
Real-world deployment
On the 60-second task, the controller trains at trial 30, verification adopts the candidate, and held-out success doubles from 35% to 70%.

      
SHARED EXPERIENCE · 1–3026.7%
8 / 30 successful trials
CANDIDATE EVALUATION · 31–4040%
4 / 10 successful trials
AFTER ADOPTION · 41–5070%
7 / 10 successes · Frozen: 5 / 10

      
BEFORE TRAINING · TRIAL 3060-second budget exhausted
AFTER ADOPTION · TRIAL 50Task completed

      
All 50 deployment trials
Both methods share trials 1–30. The same rolling window is used for both curves.
— Robo-COP– – Frozen policy

      
Training and verification record
TRAIN · AFTER TRIAL 30
Selected verification skills
30 curated episodes · 135 skill records
pick up the pot lid and put it on the grey towel
pick up the red tomato and put it on the white plate
pick up the green lime and put it on the white plate
ADOPT · AFTER TRIAL 40
The recorded verification returns CONFIRM. The adopted policy serves from trial 41.

      
Across all 50 deployment trials: Robo-COP 19/50 (38%); frozen-policy control 16/50 (32%).
Held-out evaluation. Across the three tasks, Robo-COP raises mean held-out success from 38.3% to 50.0%. Paper Section 4.3.

    


    

      
Cup recycling / Trial 72
Failure recovery
The visual monitor reports that the cup is not secure. The harness writes another program to regrasp it and continues within the same episode.

      
TRIAL 72Task completed after regrasping
01
Grasp check fails. The cup is not secured after the initial policy call.
02
Monitor stops the action. It requests revised code to close the gripper, verify the grasp, and lift.
03
Code is revised. The next program includes a grasp check and another attempt.
04
Task succeeds. The episode ends with the cup placed in the bin.


    


    
Real-world task videos
Successful and unsuccessful attempts on three real-world tasks.

      

        
TASK 01  40 S
Pot and tape placement
Put both the pot and tape on the towel.
Trial 33 · task not completed

        
TASK 02  60 S
Lid removal and object transfer
Set down the lid, then move the pot contents to the white plate.
Trial 48 · task not completed

        
TASK 03  90 S
Mustard and vegetable sorting
Put the mustard in the bowl and the vegetables in the pots.
Trial 13 · task not completed

      

    


    
Citation
@article{zhang2026robocop,
  title={Co-Evolving Robot Orchestrators and Policies through Deployment},
  author={Zhang, Xilun and Wang, Maggie and Bauer, Erik and Yu, Hong-Xing
          and Huang, Huang and Wu, Jiajun and Pavone, Marco},
  journal={arXiv preprint arXiv:2610.09228},
  year={2026},
  url={https://arxiv.org/abs/2610.09228}
}