v0.1.0

First tagged release of the PlugRL training server.

What this is

The training side of PlugRL: it holds the policy and the learning algorithm and answers environment clients over the PlugRL protocol. FPO on HalfCheetah-v5 trains CPU-only from the README quickstart.

The experiments are the substance

experiments/ holds eight completed studies and two pre-registered ones, each with its data and a findings file that states what it does not support:

• E9 ran 1 to 8 environment clients against one server: in all 12 runs the clients' step counts equalled the server's global step to the unit.
• E10 timed a VLA forward through openpi's own compiled inference entry point: 34.9 ms for the policy PlugRL ships and 100.0 ms full size, against a 1.3 ms machine boundary. PlugRL's server takes a different, uncompiled path, so those milliseconds do not transfer to it. The correction is written into the findings with its date.
• E6 is a three-seed learning curve, -315 to 1928.
• E1 disproved the hypothesis the project was originally built on, and is kept in full.

Several of these were pre-registered before any data existed, and two predictions were falsified and are reported as falsified.

Fixed on the way to a real VLA

Pi0Policy crashed on its first inference against a real pi0.5 checkpoint. openpi's image masks are numpy scalars, and batching collected them into Python lists that the tensor conversion rejects. E11's first run through the real server loop found it, and #14 fixes it with a regression test.

Known limits

• No VLA has been trained through this system yet. Pi0Policy has run evaluation end to end through the real server on LIBERO. Training is what E11 is for, and E11 is in progress.
• The Ray launcher is not a supported path.
• E3 was pre-registered and never run.