Natura-1 scaling validation
Teacher-forced action error on human and robot validation data, against hours of training data. Both axes are logarithmic, and lower error is better.
Human validation data
8 checkpoints, all used in the fitAction MSE
Log scale. Lower is better
MSE = 0.3570 · D−0.2235R² (log) = 0.981, D in hours
0.0150.020.0250.030.03550k100k200k500k1M
0.03200.02990.02530.02400.02010.01810.01750.0168
Training data, hours (log scale)
Human validation data, teacher forcing
- Training data
- Action MSE
Action MAE
Log scale. Lower is better
MAE = 0.5281 · D−0.1622R² (log) = 0.981, D in hours
0.050.060.070.080.090.150k100k200k500k1M
0.09160.08740.07820.07440.06550.06070.05910.0574
Training data, hours (log scale)
Human validation data, teacher forcing
- Training data
- Action MAE
Robot validation data
8 checkpoints, all used in the fitAction MSE
Log scale. Lower is better
MSE = 1.7193 · D−0.3893R² (log) = 0.948, D in hours
0.0080.010.0150.020.0350k100k200k500k1M
0.02720.02280.01870.01330.01100.00960.00920.0088
Training data, hours (log scale)
Robot validation data, teacher forcing
- Training data
- Action MSE
Action MAE
Log scale. Lower is better
MAE = 1.2383 · D−0.2330R² (log) = 0.964, D in hours
0.050.060.070.080.090.150k100k200k500k1M
0.10340.09210.08200.06910.06060.05520.05380.0525
Training data, hours (log scale)
Robot validation data, teacher forcing
- Training data
- Action MAE