---
license: cc-by-nc-sa-4.0
library_name: rsl-rl
tags:
  - reinforcement-learning
  - robotics
  - humanoid
  - isaac-lab
  - rsl-rl
  - locomotion
  - pytorch
  - onnx
pipeline_tag: reinforcement-learning
---

# Dropbear Locomotion v0.2.0

RSL-RL PPO checkpoint and deployment exports for COM-guided Dropbear humanoid
locomotion in NVIDIA Isaac Lab.

![Actual training environment in the live viewer](media/dropbear_train_live_domain_randomization.png)

## Training lineage and credits

The original Dropbear policy weights through approximately training iteration
7000 are credited to [Hyperspawn Robotics](https://github.com/Hyperspawn).
The subsequent continuation layers—including COM control, gait/contact reward
shaping, checkpoint adapters, evaluation, and viewer integration—are
co-authored with [robit-man](https://github.com/robit-man). Iteration numbers
are retained to make the inherited and continued weight lineage explicit.

## Results

### Experimental Stage 65b terrain branch

`model_10479.pt` uses 162 actor and 176 critic inputs, including a 104-value
terrain-height scan. It was trained at `+0.20 m/s` forward on 35% flat and 65%
anchored-box terrain. At the final update it reported 84.46% normal time-outs,
15.54% bad-orientation terminations, zero feet-overlap terminations, a 736.2
step mean episode, and mean terrain level 2.57.

The branch is useful for continued obstacle and gait research but is not
claimed converged. Reward and horizon were still varying over the last 20
updates.

![Stage 65b training metrics](media/stage65b_obstacle_training_metrics.png)

### Validated v0.1.0 flat-ground checkpoint

The frozen `model_8745.pt` completed 128 deterministic 20-second cardinal
trials at 0.20 m/s with zero falls:

| Direction | Trials | Mean achieved COM velocity | Mean error |
|---|---:|---:|---:|
| Forward | 32 | `(+0.209, +0.014)` m/s | 0.094 m/s |
| Backward | 32 | `(-0.148, +0.005)` m/s | 0.119 m/s |
| Left | 32 | `(+0.038, +0.163)` m/s | 0.097 m/s |
| Right | 32 | `(+0.004, -0.119)` m/s | 0.149 m/s |

Conditions: plane, no pushes, nominal policy-joint reset, 1,000 policy steps,
strict foot diagnostics at 0.16 m center distance and 0.10 m anatomical
lateral separation. Machine-readable reports are in `evaluation/`.

## Files

- `checkpoints/dropbear_locomotion_v0.2.0/model_10479.pt`: experimental
  terrain-scan/low-obstacle continuation checkpoint.
- `checkpoints/dropbear_locomotion_v0.2.0/terrain_scan_branch_parent_model_10280.pt`:
  exact adapted parent for the terrain branch.
- `configs/release_v0.2.0/`: resolved Stage 65b configuration.
- `training/stage65b/`: Stage 65b TensorBoard history.
- `checkpoints/dropbear_locomotion_v0.1.0/model_8745.pt`: full checkpoint.
- `checkpoints/dropbear_locomotion_v0.1.0/policy.pt`: TorchScript actor.
- `checkpoints/dropbear_locomotion_v0.1.0/policy.onnx` plus `.onnx.data`: ONNX actor.
- `checkpoints/dropbear_locomotion_v0.1.0/warm_start_model_8666.pt`: parent checkpoint.
- `dropbear_walk/isaaclab_asset/dropbear.usd`: robot USD.
- `configs/release_v0.1.0/`: resolved training configuration.

The v0.2 actor consumes 162 values and emits 14 normalized joint-position
actions at 50 Hz. The v0.1 actor consumes 58 values. See
`docs/CHECKPOINTS.md` before integrating either policy.

## Intended use and limitations

This release is intended for simulation research, reproducibility, continued
policy training, and deployment-tooling development. It is not a physical
robot safety controller. Only v0.1.0 has the reported deterministic
flat-ground validation. The v0.2.0 terrain branch, pushes, speeds above
0.20 m/s, crouched locomotion, and floor recovery are outside the validated
boundary.

Source and full documentation:
https://github.com/robit-man/dropbear-locomotion