RACER: Rich Language-Guided Failure Recovery Policies for Imitation Learning

*Equal Contribution
1Department of Computer Science and Engineering, University of Michigan, 2Robotics Department, University of Michigan

TL;DR: Visuomotor policies trained with rich language instructions and failure recovery behaviors demonstrate superior robustness and adaptability.

Rich language instructions provide more comprehensive details for failure recovery, such as failure analysis, spatial movements, target object attributes, and the expected outcome, and can guide the policy with more accurate control while serving as a form of regularization to prevent overfitting and improve generalization.




Abstract

Developing robust and correctable visuomotor policies for robotic manipulation is challenging due to the lack of self-recovery mechanisms from failures and the limitations of simple language instructions in guiding robot actions. To address these issues, we propose a scalable data generation pipeline that automatically augments expert demonstrations with failure recovery trajectories and fine-grained language annotations for training. We then introduce Rich languAge-guided failure reCovERy (RACER), a supervisor-actor framework, which combines failure recovery data with rich language descriptions to enhance robot control. RACER features a vision-language model (VLM) that acts as an online supervisor, providing detailed language guidance for error correction and task execution, and a language-conditioned visuomotor policy as an actor to predict the next actions. Our experimental results show that RACER outperforms the state-of-the-art Robotic View Transformer (RVT) on RLbench across various evaluation settings, including standard long-horizon tasks, dynamic goal-change tasks and zero-shot unseen tasks, achieving superior performance in both simulated and real world environments.

Model Architecture

RACER is a flexible supervisor-actor framework that enhances robotic manipulation through language guidance for failure recovery.


RACER Framework Overview

Data Augmentation Pipeline

We propose a scalable rich language-guided failure recovery data augmentation pipeline to collect new trajectories from expert demos.



We first generate large data in simulation, then collect small real-world data for sim2real transfer.

Simulation Data (18 tasks, 10,159 trajectories)

Real World Data (4 tasks, 180 trajectories)

Experimental Results


Real World Videos (2X default speed)

We evaluate the RVT baseline, RACER-scratch (RACER trained with rich instructions on real-world data only), RACER-simple (RACER trained with simple instructions on both simulated and real-world data), and RACER-rich (RACER trained with rich instructions on both simulated and real-world data).

(→ means there is task goal online change to test model robustness.)


Task: Open Drawer

Task Goal: Open the (top → bottom → top) drawer.

Failure: RVT

Failure: RACER-scratch

Failure: RACER-simple

Success: RACER-rich

Task: Put Item on Shelf

Task Goal: Put the Oreo on the top shelf.

Failure: RVT

Failure: RACER-scratch

Failure: RACER-simple

Success: RACER-rich

Task: Push Buttons

Task Goal: Push the blue button, then the red button.

Failure: RVT

Failure: RACER-scratch

Failure: RACER-simple

Success: RACER-rich

Task: Pick and Place Fruit

Task Goal: Put the orange into the (red → white) bowl.

Failure: RVT

Failure: RACER-scratch

Failure: RACER-simple

Success: RACER-rich

Simulation Videos

Failure Recovery with RACER

We test RACER to see if it can recover from RVT failures or self-failures in regular RLbench tasks.

Task Goal: Close the navy jar.

Failure: RVT

Success: RACER

Task Goal: Stack the wine bottle to the left rack.

Failure: RVT

Success: RACER

Task Goal: Slide block to the green target.

Failure: RVT

Success: RACER

Task Goal: Stack four purple blocks.

Failure: RVT

Success: RACER

Task Goal Online Change Evaluation

We introduce a novel setting to assess the model's robustness by deliberately switching the task goal during execution.

Task Goal: Close the (olive → cyan) jar.

Failure: RVT

Success: RACER

Task Goal: Screw in the (gray → purple) light bulb.

Failure: RVT

Success: RACER

Task Goal: Open the (middle → bottom) drawer.

Failure: RVT

Success: RACER

Task Goal: Push the (maroon → green) button, then push the (green → maroon) button.

Failure: RVT

Success: RACER

Unseen Task Evaluation

We evaluate RACER on new tasks to examine its zero-shot adaptability.

Task Goal: Close the bottom drawer.

Failure: RVT

Success: RACER

Task Goal: Move the block onto the green target.

Failure: RVT

Success: RACER

Task Goal: Pick up the yellow cup.

Failure: RVT

Success: RACER

Task Goal: Reach the red target.

Failure: RVT

Success: RACER

RACER with Human Language Intervention

RACER + H indicates that humans can decide to modify VLM instructions as needed. Our RACER, even though trained on rich instructions, is able to effectively understand and follow unseen human commands to enable better failure recovery when the VLM can not lead to task success.

Task Goal: Put the cylinder in the shape sorter.

Failure: RACER

Success: RACER + H

Task Goal: Put the money in the bottom shelf.

Failure: RACER

Success: RACER + H

Task Goal: Sweep dirt to the short dustpan.

Failure: RACER

Success: RACER + H

Task Goal: Stack the other cups on top of the maroon cup.

Failure: RACER

Success: RACER + H

Failure Cases

Due to the limited visual understanding capabilities, the current Vision-Language Model (fine-tuned LLaVA-LLaMA3-8B in our experiments) occasionally fails to accurately recognize task failures, leading to the generation of hallucinated instructions. In addition, the visuomotor policy sometimes may also predict inappropriate actions.

Task Goal: Open the top drawer.

Failure: RACER

Task Goal: Put the lemon into the red bowl.

Failure: RACER

Task Goal: Push the (blue → green) button.

Failure: RACER

Task Goal: Put the Oreo on the top shelf.

Failure: RACER

BibTeX

@misc{dai2024racerrichlanguageguidedfailure,
      title={RACER: Rich Language-Guided Failure Recovery Policies for Imitation Learning}, 
      author={Yinpei Dai and Jayjun Lee and Nima Fazeli and Joyce Chai},
      year={2024},
      eprint={2409.14674},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2409.14674}, 
}