Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Improving the performance of AI-powered Affordable Robotics for Assistive Tasks

Dharunish Yugeswardeenoo

TL;DR

The paper tackles the high demand and accessibility gap in assistive robotics by proposing a low-cost, AI-driven robotic arm trained via imitation learning. It introduces Phased Action Chunking Transformer (PACT) to inject temporal context through a phase token and Temporal Ensembling (TE) to smooth trajectories, enabling robust performance with smaller models. Empirically, the approach achieves over 90% task accuracy across feeding, wiping, and medicine-fetching tasks, with PACT enabling about a 5x model-size reduction while maintaining ~75% accuracy and TE providing smoother, safer motions. The work demonstrates a practical, open-source path to affordable assistive robotics and outlines future work toward bimanual manipulation and mobility enhancements.

Abstract

By 2050, the global demand for assistive care is expected to reach 3.5 billion people, far outpacing the availability of human caregivers. Existing robotic solutions remain expensive and require technical expertise, limiting accessibility. This work introduces a low-cost robotic arm for assistive tasks such as feeding, cleaning spills, and fetching medicine. The system uses imitation learning from demonstration videos, requiring no task-specific programming or manual labeling. The robot consists of six servo motors, dual cameras, and 3D-printed grippers. Data collection via teleoperation with a leader arm yielded 50,000 video frames across the three tasks. A novel Phased Action Chunking Transformer (PACT) captures temporal dependencies and segments motion dynamics, while a Temporal Ensemble (TE) method refines trajectories to improve accuracy and smoothness. Evaluated across five model sizes and four architectures, with ten hours of real-world testing, the system achieved over 90% task accuracy, up to 40% higher than baselines. PACT enabled a 5x model size reduction while maintaining 75% accuracy. Saliency analysis showed reliance on key visual cues, and phase token gradients peaked at critical trajectory moments, indicating effective temporal reasoning. Future work will explore bimanual manipulation and mobility for expanded assistive capabilities.

Improving the performance of AI-powered Affordable Robotics for Assistive Tasks

TL;DR

The paper tackles the high demand and accessibility gap in assistive robotics by proposing a low-cost, AI-driven robotic arm trained via imitation learning. It introduces Phased Action Chunking Transformer (PACT) to inject temporal context through a phase token and Temporal Ensembling (TE) to smooth trajectories, enabling robust performance with smaller models. Empirically, the approach achieves over 90% task accuracy across feeding, wiping, and medicine-fetching tasks, with PACT enabling about a 5x model-size reduction while maintaining ~75% accuracy and TE providing smoother, safer motions. The work demonstrates a practical, open-source path to affordable assistive robotics and outlines future work toward bimanual manipulation and mobility enhancements.

Abstract

By 2050, the global demand for assistive care is expected to reach 3.5 billion people, far outpacing the availability of human caregivers. Existing robotic solutions remain expensive and require technical expertise, limiting accessibility. This work introduces a low-cost robotic arm for assistive tasks such as feeding, cleaning spills, and fetching medicine. The system uses imitation learning from demonstration videos, requiring no task-specific programming or manual labeling. The robot consists of six servo motors, dual cameras, and 3D-printed grippers. Data collection via teleoperation with a leader arm yielded 50,000 video frames across the three tasks. A novel Phased Action Chunking Transformer (PACT) captures temporal dependencies and segments motion dynamics, while a Temporal Ensemble (TE) method refines trajectories to improve accuracy and smoothness. Evaluated across five model sizes and four architectures, with ten hours of real-world testing, the system achieved over 90% task accuracy, up to 40% higher than baselines. PACT enabled a 5x model size reduction while maintaining 75% accuracy. Saliency analysis showed reliance on key visual cues, and phase token gradients peaked at critical trajectory moments, indicating effective temporal reasoning. Future work will explore bimanual manipulation and mobility for expanded assistive capabilities.
Paper Structure (11 sections, 5 figures, 1 table)

This paper contains 11 sections, 5 figures, 1 table.

Table of Contents

  1. Introduction
  2. Methodology
  3. Experimental Setup
  4. Results
  5. Saliency Analysis
  6. CONCLUSION
  7. Appendix
  8. PACT Architecture
  9. Temporal Ensembling
  10. Image Saliency
  11. PACT Saliency

Figures (5)

  • Figure 1: Three different assistive tasks and respective sub-tasks.
  • Figure 2: PACT Architecture
  • Figure 3: Visual of TE's impact is shown in terms of the evolution of motor angles over a time. The 64-dimensional model without TE (left) outputs jittery and inconsistent movement. The right graph shows the same model but with TE applied, displaying smoother and slower motion. Both models are examined on the same test case.
  • Figure 4: The above figure highlights the saliency for the input images for the 512 dimensions and 32 dimensional models on a test case in the fetching medicine task.
  • Figure 5: The above figure displays the saliency of the phase token at each frame in an episode of the fetching medicine task. The model's attention towards the phase token peaks at an instance where visual ambiguity was frequently observed.