AlphaBrainGroup/alphabrain-rlt-5traj-alltasks-libero-goal

AlphaBrain-RLT (5-trajectory · LIBERO-Goal · all 10 tasks)

RL-Token (RLT) fine-tuning release from the AlphaBrain framework. Lightweight Actor–Critic on top of a frozen VLA backbone, covering all 10 LIBERO-Goal tasks in the low-data (5 trajectories / task) setting.

RL-Token compresses the frozen VLA's hidden state through an information bottleneck encoder, then trains an Actor head (residual action edits) and a Critic head (state value) with off-policy reinforcement learning. The VLA backbone — AlphaBrainGroup/qwenoft-5traj-libero-goal — is frozen during RL, so only a few hundred MB of parameters are learned on top.

Overview

Architecture	RL-Token (encoder + Actor + Critic) on top of a frozen VLA
Base VLA (frozen)	AlphaBrainGroup/qwenoft-5traj-libero-goal
Encoder	Transformer bottleneck, dim=256, 2 layers, 4 heads
Actor	hidden_dim=512, ref_dropout=0.5, fixed_std=0.1
RL algorithm	TD3-style off-policy Actor–Critic
Task scope	LIBERO-Goal — all 10 tasks trained jointly
Iterations released	iter_00400 (eval default)

Results

Evaluated with 50 episodes per task across all 10 LIBERO-Goal tasks (eval release 2026-04-16):

Task	Success rate
open the middle drawer of the cabinet	100 %
put the bowl on the stove	100 %
put the wine bottle on top of the cabinet	94 %
open the top drawer and put the bowl in	68 %
put the bowl on top of the cabinet	100 %
push the plate to the front of the stove	86 %
put the cream cheese in the bowl	80 %
turn on the stove	100 %
put the bowl on the plate	100 %
put the wine bottle on the rack	92 %
Overall	92 %

Numbers are from our internal release eval; per-run variance is a few percentage points. Reproduction numbers higher or lower than reported are expected — please file an issue / PR with details.

Files

├── README.md                           model card
├── metrics.json                        RL training metrics trace
├── eval_summary.json                   per-task SR + overall (2026-04-16 release)
└── rl_offpolicy_iter_00400/
    ├── actor.pt                        Actor (~1.8 MB)
    ├── critic.pt                       Critic (~3.3 MB)
    └── encoder.pt                      Bottleneck encoder (~517 MB)

Usage

This checkpoint is evaluated via a custom RL eval pipeline. Clone the AlphaBrain framework and download both the base VLA and this RLT release:

git clone https://github.com/AlphaBrainGroup/AlphaBrain.git
cd VLA-Engine-Developer
pip install -e .

export PRETRAINED_MODELS_DIR=/path/to/models
export LIBERO_PYTHON=/path/to/envs/libero/bin/python   # env with robosuite
export LIBERO_HOME=/path/to/LIBERO

# Download base VLA
huggingface-cli download AlphaBrainGroup/qwenoft-5traj-libero-goal \
    --local-dir results/training/QwenOFT-5traj-libero_goal/final_model

# Download RLT release
huggingface-cli download AlphaBrainGroup/alphabrain-rlt-5traj-alltasks-libero-goal \
    --local-dir results/rlt_training_TD3/rlt_5traj_alltasks_release/rl_offpolicy/checkpoints

# Run the 10-task LIBERO-Goal evaluation (splits across 3 GPUs)
bash scripts/run_rl_scripts/run_eval_rlt.sh \
    results/rlt_training_TD3/rlt_5traj_alltasks_release/rl_offpolicy \
    0,1,2

The eval script:

1. Loads the base VLA + Actor / Critic / Encoder
2. Splits the 10 LIBERO-Goal tasks across the 3 GPUs you specify
3. Rolls out 50 episodes per task
4. Writes per-shard JSONs and an aggregated summary.json

Reproduction

Training is two-phase — (1) supervised VLA fine-tuning on 5 trajectories per task, (2) RL-Token Actor–Critic fine-tuning with a frozen VLA. The full pipeline can be reproduced with:

bash scripts/run_rl_scripts/run_rlt_5traj_alltasks.sh

See the RL scripts in the framework for details, required env variables, and per-phase configuration.

License

MIT — see the parent repository.

Citation

@misc{alphabrain2026,
  title  = {AlphaBrain: A Modular Open-Source Framework for Embodied Intelligence Research},
  author = {AlphaBrain Team},
  year   = {2026},
  url    = {https://github.com/AlphaBrainGroup/AlphaBrain}
}