Partial Motion Imitation for Learning Cart Pushing with Legged Manipulators

Dimensions overall score 7.0

• The proposed partial imitation learning approach transfers locomotion style from a simpler task to cart loco-manipulation.method 95%
  Evidencepartial

  This work proposes a partial imitation learning approach that transfers the locomotion style learned from a locomotion task to cart loco-manipulation.

  Implicationpartial

  This is a core methodological claim stated in the abstract and elaborated upon in the introduction.

  Verificationpartial

  partial

• The learned policy successfully pushes a cart along diverse trajectories in IsaacLab and transfers effectively to MuJoCo.result 95%
  Evidencepartial

  We conduct experiments demonstrating that the learned policy successfully pushes a cart along diverse trajectories in IsaacLab and transfers effectively to MuJoCo.

  Implicationpartial

  This is a key result claim stated in the abstract and supported by the experimental results in Table V.

  Verificationpartial

  partial

• The proposed approach achieves more stable and accurate loco-manipulation behaviors compared to baselines.result 95%
  Evidencepartial

  We also compare our method to several baselines and show that the proposed approach achieves more stable and accurate loco-manipulation behaviors.

  Implicationpartial

  This is a comparative result claim stated in the abstract and supported by the quantitative data in Table V.

  Verificationpartial

  partial

• The legged manipulator platform consists of a WidowX AI arm mounted on a Unitree Go2 quadruped, tasked with pushing a kid-sized shopping cart.technical 90%
  Evidencepartial

  Our platform is a legged manipulator consisting of a WidowX AI arm [39] mounted on a Unitree Go2 quadruped [40], tasked with pushing a kid-sized shopping cart [41].

  Implicationpartial

  This describes the specific technical setup and task, which is crucial for understanding the experiments.

  Verificationpartial

  partial

• Extensive domain and terrain randomization were used to train a robust locomotion policy.method 90%
  Evidencepartial

  A robust locomotion policy is first trained with extensive domain and terrain randomization...

  Implicationpartial

  This details a specific aspect of the training methodology that contributes to robustness.

  Verificationpartial

  partial

• The Partial AMP method achieves a survival rate of 99.9% in IsaacLab, significantly higher than 'No AMP' (96.8%) and 'Full AMP' (97.2%).result 95%
  Evidencepartial

  Partial AMP (ours)99.9±0.10.045±0.000 0.085±0.001 0.035±0.001 0.038±0.0000.293±0.001

  Implicationpartial

  This is a specific, verifiable quantitative result comparing the proposed method to baselines.

  Verificationpartial

  partial

• Full AMP tends to overfit to arm states, resulting in unstable contact behavior.limitation 85%
  Evidencepartial

  Full AMP tends to overfit to the arm states observed during training, often keeping the arm close to the base, resulting in unstable contact behavior

  Implicationpartial

  This is a specific limitation identified for a baseline method, explaining a failure mode.

  Verificationpartial

  partial

Competitive landscape

Enables legged robots to push carts effectively by transferring learned locomotion styles to manipulation tasks.