PYTHON · NUMPY

Run the Released Checkpoint

The released checkpoint has 17,923 parameters. What does running its exact forward pass demonstrate?

Run the exact forward pass of the released checkpoint, understand its 17,923 parameters, and hand off to the live demo.

01Challenge

Implement the released policy's forward pass (z-score in, three layers, z-score out) and confirm it emits a valid 3-D joint delta.

Predict firstThe released checkpoint has 17,923 parameters. What does running its exact forward pass demonstrate?

02Model

Everything you built is the released checkpoint in miniature. forge-arm-reach-bc is a 7 to 128 to 128 to 3 network with SiLU activations and z-scored inputs and outputs, 17,923 parameters, 72 kilobytes. Here you run its exact forward pass. On the checkpoint page, this same code runs the real published weights and closes the loop on the arm live, no server. You have gone from demonstrations to a policy you can run; this is where it becomes real.

03Guided practice

1. The forward pass. Z-score the observation, apply three linear layers with SiLU between them, then un-z-score the output. That is the whole policy.

2. The shape. Weights of shape 7→128, 128→128, 128→3. Count them and you land near the checkpoint's 17,923.

3. Run it for real. Open /research/checkpoint to watch the actual released weights drive the same MuJoCo arm they were trained on, and record your own dataset in Forge to train a policy of your own.

04Feedback

PASS WHEN PASS when the forward pass runs and returns a single 3-D joint-delta action.

If your run fails, check:

A cell errored, run them in order.
The policy must output exactly three numbers, one delta per joint.

05Retrieve & space

What is the released checkpoint's architecture? (7 → 128 → 128 → 3, SiLU, z-scored I/O).
How many parameters does it carry? (17,923).
What runs the real weights live? (the checkpoint page at /research/checkpoint).

06Mastery & project

You can run the released policy's forward pass and connect it to the live on-device demo.

The full arc, in your hands: demonstrations → dataset → trained policy → a released checkpoint you can run on the page.

GO FURTHER · THE LIVE PLAYGROUND

Run the released checkpoint live ↗ · the exact forward pass, on-device in your browser.

← Beyond Frozen: The On-Robot Flywheel How Many Seeds Does a Ranking Need? →

PYTHON · NUMPY

Run the Released Checkpoint

Run the released checkpoint's exact forward-pass architecture and produce a joint-delta action.

You read

the arrays and values already in scope

You change

the code you write in each cell

Fixed

the dataset and the checks that grade you