VLA Robot Candidates for Flower Trimming and Bouquet Making

Summary

Flower trimming and bouquet making are harder than simple pick-and-place. The task includes aligning stem lengths, removing leaves, orienting flowers, holding a bundle with one hand while manipulating wrap or string with the other, and potentially using tools. The main buying criteria are bimanual manipulation, workspace reach, swappable grippers/hands, data collection pipeline, and customization headroom.

Priority	Candidate	First judgment
1	OpenArm	Strongest combination of official LeRobot support, 7-DoF bimanual arms, low cost, and open-source design.
2	ALOHA / ALOHA 2	Strongest ACT/VLA research reproduction and bimanual data collection ecosystem.
3	Galaxea R1 Pro / A1 dual	Purchasable high-performance bimanual/mobile option with force-control advantages.
4	Reachy 2	Official LeRobot support, VR teleoperation, and humanoid bimanual form factor.
5	Dual xArm 7	Good reach and payload for the price, but the bimanual VLA stack needs custom integration.
6	Dual Franka Research 3	Excellent force control and research ecosystem, with high budget and integration cost.
7	Unitree G1 EDU	Interesting LeRobot-supported humanoid candidate, but walking and safety complexity is high for table flower work.

The practical path is to first validate the LeRobot record/train/inference loop with SO-101 or LeKiwi, then select OpenArm, ALOHA 2, Galaxea R1 Pro, or Reachy 2 as the main experiment platform.

Source Basis for This Update

The shortlist is not based on manufacturer copy alone. It is cross-checked against these sources:

Source	How it was used
Hugging Face LeRobot official docs	Filter robots that are easier to put into record/train/inference workflows
Humanoid Guide / HumanoidSpecs hands directory	Broad recent scan of hands and grippers
Robot Manipulation hardware list	Wider research-manipulation hardware context
Manufacturer/lab official docs	Spec and integration checks for WUJI, ORCA, SharpaWave, and North
Papers and preprints	Evidence for open-source reproducibility, especially ORCA

Evaluation Criteria

Criterion	Why it matters
Purchasability	Demo-only robots are excluded because the goal is to train and deploy a custom VLA.
LeRobot support	It shortens the path to recording, training, Hub sharing, and policy deployment.
Bimanual manipulation	Bouquet work often needs one hand to hold while the other aligns, ties, wraps, or uses tools.
Reach/payload	The robot must cover flowers, scissors, wrapping paper, vases, and trays on a tabletop.
Dexterity	Contact surface, force control, and end-effector options matter for avoiding damage.
Durability	Repeated data collection stresses motors, gears, frames, and cables.
Open-source/customization	Cameras, custom grippers, safety interlocks, and VLA inference need direct integration.
Price	Cameras, compute, grippers, safety cells, and spare parts add to the robot cost.

Official LeRobot Candidates

Candidate	Purchase/build path	Bimanual fit	Flower-task fit	Main risk
SO-101	Very good low-cost DIY/kit path	Single or leader-follower pair	Entry point for data pipeline	Reach, torque, and durability are limited
Koch v1.1	Good Dynamixel-based DIY/kit path	Dual setup possible	Low-cost teleop practice	Payload and stiffness are limited
LeKiwi	Good SO-101 plus mobile base path	Single arm by default	Mobile manipulation pipeline validation	No bimanual coordination
OpenArm	Good open-source BOM path	Official bimanual type support	First-choice main experiment candidate	Young ecosystem and supply/build risk
Reachy 2	Quoted, high-cost purchase	Bimanual plus mobile base	Strong humanoid bouquet-work candidate	Price, procurement, and gripper dexterity
Unitree G1	Basic/EDU purchasable	Humanoid bimanual arms	Humanoid VLA demo candidate	EDU access, safety, and shorter arm reach

Main Platform Comparison

Candidate	Reach/payload	Dexterity	Price band	Open/custom	VLA fit	Judgment
OpenArm	7-DoF, 4.1kg nominal/6kg peak per arm	Needs end-effector selection	About $6.5K bimanual BOM	Very high	Official LeRobot support	First candidate to verify procurement for
ALOHA 2	ViperX-based, about 750mm reach/light payload	Default and improved grippers	About $20K+	High	ACT/VLA research standard	Strong for paper reproduction and live content
Galaxea R1 Pro	Dual 7-DoF arms, kg-class per arm	G1 gripper/force control	About $70K for R1 Pro	Medium-high	Native VLA/ROS ecosystem	Saves time at higher cost
Reachy 2	Dual 7-DoF arms, 3kg per arm	Default gripper may need replacement	High, quoted	Open-source software	Official LeRobot support	Strong humanoid UX and VR teleop
Dual xArm 7	700mm/3.5kg	Wide gripper options	About $22K+ for two arms	Good SDK/ROS	Custom integration required	Robust cell candidate for the price
Dual Franka Research 3	855mm/3kg with joint torque sensing	Excellent force control	Very high	Strong research ecosystem	Very high data quality	Premium candidate if budget allows
Dual UR5e/UR10e	850-1300mm / 5-12.5kg	Many industrial grippers	Used/quoted	Strong ROS2	Custom integration required	Durable but needs extra sensing for delicacy

Hand and Gripper Candidates

Candidate	Character	Recommended use
OnRobot Soft Gripper	Silicone soft gripper	Direct flower grasping and wrapping-paper holding
Robotiq 2F/3F	Industrial parallel/3-finger grippers	Holding stem bundles, tools, and fixtures
LEAP Hand	16-DoF open-source dexterous hand	In-hand alignment and dexterity research
WUJI Hand	20 active DoF, ROS 2/SDK, around 580 g	High-DoF experiments on stems, tools, and wrapping paper
ORCA Hand	17 DoF open-source hand, Touch line	Fingertip modification and tactile datasets
SharpaWave	22 active DoF and high-resolution tactile-array claims	Petal/stem contact classification and force-cap research
Allegro Hand	16-DoF commercial research hand	Higher-end dexterous manipulation
Ability Hand	Prosthetic hand with touch sensing	Reducing damage to petals and stems

The pragmatic first test is not an expensive dexterous hand. Mount OnRobot Soft Gripper or Robotiq 2F with replaceable silicone pads on OpenArm/ALOHA/xArm and measure flower damage. Move to LEAP Hand, WUJI Hand, ORCA Hand, SharpaWave, Ability Hand, or Allegro Hand only if in-hand rotation, tactile force caps, or fine orientation becomes necessary.

Watch-Only Candidate

Candidate	Judgment
Sharpa North	Reference platform for SharpaWave hands on a wheeled full-body robot. Keep it as a watch candidate until SDK, pricing, lead time, and safety access are public enough to evaluate.

Buying Scenarios

Budget-Constrained

Validate LeRobot record/train/inference with SO-101 or LeKiwi.
Source or build bimanual OpenArm.
Add silicone pads and simple force/pressure sensing to the base gripper.
Collect data in stages: single flower pick, stem alignment, then simple tying.

Research-Reproduction

Buy or build ALOHA 2 / Trossen ALOHA.
Reproduce ACT/ALOHA baselines.
Fine-tune OpenVLA, SmolVLA, or pi0-style policies.
Define bouquet making as an ALOHA-style benchmark.

Time-Saving

Request quotes for Galaxea R1 Pro or Reachy 2.
Collect human demonstrations through VR/isomorphic teleoperation.
Connect LeRobot or a custom ROS2 data conversion pipeline.
Validate VLA inference inside a safety cell.

Archive Version

The longer live-content research report is also available in the Archive version.