Recoleta
GitHub
Source note

VILAS: A VLA-Integrated Low-cost Architecture with Soft Grasping for Robotic Manipulation

Authors
Zijian An, Hadi Khezam, Bill Cai, Ran Yang, Shijie Geng, Yiming Feng, et al.
Source
arXiv
Published
Collection
Embodied AI
Vision Language ActionRobot ManipulationLow Cost RoboticsSoft GraspingTeleoperation Data
Open arXivMarkdown source

Summary

VILAS is an $8,000 robotic manipulation platform for fine-tuning and deploying VLA policies on fragile-object grasping. It combines an industrial cobot, dual RGB-D cameras, teleoperation data collection, and a kirigami soft gripper extension.

Problem

  • Fragile objects such as grapes can be crushed by rigid grippers or abrupt robot motion, so manipulation needs gentle contact and stable grasping.
  • Many VLA robot setups depend on expensive integrated hardware; the paper compares VILAS at about $8,000 against an ALOHA kit at about $33,000.
  • Force sensors, tactile arrays, and force-control systems add cost and integration work, which limits access for VLA policy research.

Approach

  • The platform uses a Fairino FR5 arm, Jodell RG52-50 gripper, GELLO teleoperation arm, RealSense D455 base camera, RealSense D405 wrist camera, and a ZMQ-based control stack.
  • A human operator collects demonstrations with GELLO. Each timestep stores a 7-DoF joint state, two RGB camera views, and a language prompt.
  • A PEBA kirigami extension mounts on the parallel gripper and bends under compression, spreading contact over the fruit and reducing the need for force sensing.
  • The authors fine-tune pi_0, pi_0.5, and GR00T N1.6 from public checkpoints on the same dataset: 100 episodes, 1,200 frames per episode, trained for 50,000 iterations on an NVIDIA H200.
  • During deployment, the robot runs at 20 Hz. pi_0 and pi_0.5 output 50-action chunks, while GR00T N1.6 outputs 16-action chunks.

Results

  • Total hardware cost is about $8,000, compared with about $33,000 for ALOHA.
  • On the grape task, each model was tested over 50 trials. Single-grasp success was 70% for pi_0, 84% for pi_0.5, and 82% for GR00T N1.6.
  • Multi-grasp success, defined as at least two consecutive successful grasps, was 22% for pi_0, 36% for pi_0.5, and 58% for GR00T N1.6.
  • Mean inference latency was 73.8 ms for pi_0, 82.8 ms for pi_0.5, and 63.6 ms for GR00T N1.6.
  • Per-step inference cost was 1.48 ms for pi_0, 1.66 ms for pi_0.5, and 3.98 ms for GR00T N1.6, reflecting the shorter action horizon for GR00T N1.6.
  • The strongest claim is system-level: current VLA policies can be fine-tuned and deployed for delicate real-world grasping on a lower-cost modular robot, with GR00T N1.6 giving the best sequential grasp reliability in this setup.

Link