TontonTremblay/jaco-pick-experiment-tmp

Jaco vs Franka Pick Task Comparison

Experiment Setup

• Scene: iTHOR house 1
• Seed: 42
• Episodes: 2 per robot
• Task: Pick (pick up object)

Results

Robot	Success Rate	Episodes	Notes
Franka (Droid)	100%	2/2	Works reliably
Jaco	0%	0/1	IK failures during path execution

Key Finding

The Jaco robot fails due to IK failures during interpolated path following, NOT because of:

• Grasp selection (grasps are the same for both robots)
• Initial IK feasibility check (passes)
• Final pose reachability (endpoints are feasible)

The issue is that Jaco's IK solver cannot reliably track all intermediate poses along the motion path. Franka succeeds because:

1. It has a larger workspace
2. Its IK solver handles path interpolation better

Recommendations

1. Use cuRobo for motion planning instead of IK-based path following
2. Generate Jaco-specific grasps that are more conservative
3. Reduce motion speed further to give IK more time
4. Consider trajectory optimization instead of linear interpolation

Files

• franka_baseline/ - Franka experiment (100% success)
• jaco_comparison/ - Jaco experiment (0% success)
• JACO_IK_INVESTIGATION.md - Detailed analysis
• jaco_datagen_configs.py - Jaco configuration with tuned parameters