Final policy checkpoints produced by reproducing RAMBO: RL-augmented Model-based Optimal Control for Whole-body Loco-manipulation.

Each model_*.pt is a CRL2 PyTorch training checkpoint. It contains the policy, value network, optimizer state, observation normalizer, and completed iteration count. It is not a Transformers model or an inference-only safetensors file.

The accompanying agent.yaml and env.yaml files record the training configuration. The checkpoint paths intentionally exclude training videos, TensorBoard events, W&B files, and Python pickle configuration files.

SHA-256 checksums:

Use the matching RAMBO/Isaac Lab environment and place a downloaded checkpoint under the corresponding CRL2 log root before launching play.py. For example, for the Quadruped policy:

For Biped, substitute task Isaac-RAMBO-Biped-Go2-v0, log root rambo_biped, run name hf_biped, and checkpoint 4000.

These checkpoints were generated from source commit 6f6c87c614255a5ada063da40206a2fa67054cb6 of catachiii/rambo, using Isaac Sim 4.5.0, Isaac Lab, CRL2, and PyTorch. The local reproduction includes checkpoint resume/normalizer-state fixes; use the accompanying workspace source for the most reliable loading behavior.

env.yaml is an Isaac Lab configuration dump that contains NumPy-specific YAML tags. Treat it as trusted project configuration and load it only within a trusted RAMBO/Isaac Lab environment.

The source RAMBO repository is licensed under CC BY-NC 4.0. These artifacts are shared under the same non-commercial terms; see LICENSE-CC-BY-NC-4.0.md.

If you use RAMBO, please cite:

PyTorch .pt files are pickle-based checkpoints. Only load files obtained from trusted sources, and use them in simulation before considering any hardware deployment.