Bittle doing something cute!
- Had a meeting with Harvard PhD student about their research.
- Started setup of IsaacSim
Meeting Notes
Expert level learning, provide expert level policy and provide a decision controller to learn the policy. Only learning from the expert demonstration using decision transformers. Further shrink it (a lot to improve the pruning). Deriving an expert level policy just from an expert level demonstration. Similar to how they trained chatGPT. The expert chooses between the two options. Then putting it on the hardware. Must need raspberry pi. The policy that it’s learning is the exact angles. They used IsaacSim.
Possible Directions
-
Using multiple inputs to sense the angles and such before sending to the policy (Learn sensing → predicting)
-
PTO isn’t the optimal way. Robot was positioning itself in a slightly tilted position. Especially for quadruple locomotion.
Outstanding Questions
-
Is there a way to modify the target to make a single heading.
-
Is there a similar way to do the imitation learning with the sensor model based approach? (Low resource learning)
-
Angle and then forward instead of one target?
What to do
- Cite all the papers and put two to three bullet points.
- Save the failures too.
Harvard Project Resources