---
license: other
library_name: rl-games
pipeline_tag: reinforcement-learning
tags:
  - reinforcement-learning
  - robotics
  - isaac-lab
  - isaac-sim
  - factory
  - ppo
  - rl-games
  - operational-space-control
  - imitation-learning
---

# Factory Feed-Forward PPO+OSC Teachers

This repository contains three non-recurrent RL-Games PPO+OSC teacher
checkpoints for Isaac Lab Factory assembly tasks:

- `Isaac-Factory-PegInsert-Direct-v0`
- `Isaac-Factory-GearMesh-Direct-v0`
- `Isaac-Factory-NutThread-Direct-v0`

The teachers were trained from the stock Isaac Lab Factory RL-Games PPO+OSC
configuration with both actor and central-value LSTM blocks removed. They are
intended as clean teacher policies for behavior cloning or DAgger-style data
collection, where each action should be a function of the current policy
observation only.

## Architecture

- Actor: `19`D policy observation -> MLP `[512, 128, 64]`, ELU -> Gaussian
  `6`D OSC action.
- Central critic: `43`D privileged state -> MLP `[512, 128, 64]`, ELU ->
  scalar value.
- RL-Games `seq_length=1`.
- Actor `rnn=null`.
- Central-value `network.rnn=null`.

The action space is the Factory `6`D operational-space-control action. These
are not FORGE checkpoints and do not use the FORGE success-prediction action
slot.

## Checkpoints

| Task | File | Stop Epoch | Training Seed |
| --- | --- | ---: | ---: |
| PegInsert | `checkpoints/peginsert_ff_ppo_osc_ep100.pth` | `100` | `0` |
| GearMesh | `checkpoints/gearmesh_ff_ppo_osc_ep150.pth` | `150` | `0` |
| NutThread | `checkpoints/nutthread_ff_ppo_osc_ep100.pth` | `100` | `0` |

These files are raw PyTorch/RL-Games `.pth` checkpoints. Load them only in a
trusted environment.

## Evaluation

Evaluation used held-out reset seeds `100,101,102` for gates and
`100,101,102,103,104,105` for confirmation.

| Task | Gate Horizon | Gate Success | Six-Seed Success | Mean Return | Mean TTS |
| --- | ---: | ---: | ---: | ---: | ---: |
| PegInsert | `150` | `376/384` (`97.92%`) | `750/768` (`97.66%`) | `350.12` | `35.72` |
| GearMesh | `300` | `378/384` (`98.44%`) | `750/768` (`97.66%`) | `719.53` | `75.02` |
| NutThread | `450` | `378/384` (`98.44%`) | `755/768` (`98.31%`) | `834.76` | `298.12` |

All gate and confirmation summaries record `agent_runtime.is_rnn=false`,
actor `rnn=null`, central-value `network.rnn=null`, and `seq_length=1`.

The JSON summaries are included under:

- `inspect/`
- `eval/`
- `videos/*_summary.json`

## Rendered Success Videos

Each video is a one-environment rollout on seed `100` using real Factory
sensor cameras. The external and wrist camera blank-frame counts were zero.

### PegInsert

<video controls width="640">
  <source src="https://huggingface.co/eishan05/factory-ff-ppo-osc-teachers/resolve/main/videos/peginsert_success_seed100.mp4" type="video/mp4">
</video>

[Open PegInsert video](https://huggingface.co/eishan05/factory-ff-ppo-osc-teachers/resolve/main/videos/peginsert_success_seed100.mp4)

- Success: `true`
- First success step: `53`
- Frames: `150`
- Blank frames: `0/150` external, `0/150` wrist

### GearMesh

<video controls width="640">
  <source src="https://huggingface.co/eishan05/factory-ff-ppo-osc-teachers/resolve/main/videos/gearmesh_success_seed100.mp4" type="video/mp4">
</video>

[Open GearMesh video](https://huggingface.co/eishan05/factory-ff-ppo-osc-teachers/resolve/main/videos/gearmesh_success_seed100.mp4)

- Success: `true`
- First success step: `47`
- Frames: `300`
- Blank frames: `0/300` external, `0/300` wrist

### NutThread

<video controls width="640">
  <source src="https://huggingface.co/eishan05/factory-ff-ppo-osc-teachers/resolve/main/videos/nutthread_success_seed100.mp4" type="video/mp4">
</video>

[Open NutThread video](https://huggingface.co/eishan05/factory-ff-ppo-osc-teachers/resolve/main/videos/nutthread_success_seed100.mp4)

- Success: `true`
- First success step: `333`
- Frames: `450`
- Blank frames: `0/450` external, `0/450` wrist

## Loading Notes

These checkpoints must be loaded with the same feed-forward RL-Games config
used for training. In Hydra override form, remove both recurrent blocks and
force sequence length to one:

```bash
'~agent.params.network.rnn' \
'~agent.params.config.central_value_config.network.rnn' \
agent.params.config.seq_length=1
```

In the source project, the runner switch was:

```bash
PHASE5_ASSEMBRAIN_DISABLE_RNN=1
```

Example task/checkpoint pairing:

```text
Isaac-Factory-PegInsert-Direct-v0  -> checkpoints/peginsert_ff_ppo_osc_ep100.pth
Isaac-Factory-GearMesh-Direct-v0   -> checkpoints/gearmesh_ff_ppo_osc_ep150.pth
Isaac-Factory-NutThread-Direct-v0  -> checkpoints/nutthread_ff_ppo_osc_ep100.pth
```

## Scope

These models are simulation teachers. They have not been validated on real
hardware. They were trained for Factory PPO+OSC policy rollouts and later
imitation-learning data generation, not for direct deployment.