ActionMap (50 demos per task)· Iteration 2
EnvironmentDatum-corner alignment
- Episodes
- OpenVLA-OFT 30 ActionMap (50 demos per task) 30
- Total duration
- OpenVLA-OFT 11:12 ActionMap (50 demos per task) 17:01
- Mean duration
- OpenVLA-OFT 22.4 s ActionMap (50 demos per task) 34.0 s
- Success
- OpenVLA-OFT 23% ActionMap (50 demos per task) 17%
- Manipulation density
- OpenVLA-OFT Share of time manipulating, five definitionsJoints moving99.0%End-effector moving87.8%Gripper engaged0.0%Contact force > 5 N75.2%Gripper engaged or in contact75.2% ActionMap (50 demos per task) Share of time manipulating, five definitionsJoints moving97.5%End-effector moving80.6%Gripper engaged0.0%Contact force > 5 N74.2%Gripper engaged or in contact74.2%
- SPARC smoothness
- OpenVLA-OFT Spectral arc length of the motionHow directly the arm moved, read from the shape of its speed spectrum. Unitless measure ≤ -1; the closer to -1, the smoother.The measure ignores how large a movement is, so a large and a small one compare directly. Shown here as the mean over this environment's measured episodes.Cutoff10 HzAmplitude threshold5%Zero-padding16×Signalend-effector translationBalasubramanian et al., On the analysis of movement smoothness, J Neuroeng Rehabil 2015. ActionMap (50 demos per task) Spectral arc length of the motionHow directly the arm moved, read from the shape of its speed spectrum. Unitless measure ≤ -1; the closer to -1, the smoother.The measure ignores how large a movement is, so a large and a small one compare directly. Shown here as the mean over this environment's measured episodes.Cutoff10 HzAmplitude threshold5%Zero-padding16×Signalend-effector translationBalasubramanian et al., On the analysis of movement smoothness, J Neuroeng Rehabil 2015.
Episode
Held-out benchmark environment
0:000:23