CogACT-LIBERO

A CogACT-Base checkpoint jointly fine-tuned on LIBERO-Spatial, LIBERO-Object, LIBERO-Goal, and LIBERO-Long (LIBERO-10) for the Efficient VLA Benchmark. One checkpoint serves all four suites.

The model combines DINOv2/SigLIP visual encoders, Llama 2 7B, and a DiT-B action head. It takes a language instruction and one agent-view image, and predicts 16 seven-dimensional actions.

Training

Setting	Value
Initialization	CogACT/CogACT-Base
Data	Four LIBERO suites, joint training
Updates	10,000
Trainable components	Vision encoders, language model, projector, action head
Optimizer	AdamW
Learning rate	2 × 10⁻⁵, constant; no warmup
Global batch size	1,024 (16 per device, 64 devices)
Weight decay / gradient clipping	0 / 1.0
Compute precision	BF16 mixed precision
Image augmentation	Enabled
Diffusion repeats per training sample	8
Seed	42

The portable recipe is in training_config.json. Weights are stored in FP32; the benchmark runs the VLM in BF16 and the action head in FP32.

Download and use

hf download guanweifan/CogACT-LIBERO --local-dir ./models/CogACT-LIBERO

Follow the benchmark's CogACT installation guide, then run from the benchmark repository:

python -m efficient_vla_benchmark.cli evaluate \
  --config configs/experiments/cogact/baseline.yaml \
  --model-path ./models/CogACT-LIBERO \
  --libero-root third_party/benchmarks/libero \
  --output-dir ./outputs/cogact-smoke

This command runs the repository's small smoke configuration. The benchmark baseline uses DDIM with 10 denoising steps, CFG 1.5, and eight executed actions per query. Its reported aggregate success rate is 90.60% over 2,000 episodes across the four suites; see the result tables.

The export includes the checkpoint, model configuration, suite-specific action statistics, and offline tokenizer. Normalization keys are libero_spatial_no_noops, libero_object_no_noops, libero_goal_no_noops, and libero_10_no_noops. File checksums are provided in SHA256SUMS.

Attribution and license

Based on CogACT (paper). CogACT weights are released under MIT; the included Llama 2 components retain the Llama 2 Community License and Acceptable Use Policy. See LICENSE, LLAMA2_LICENSE.txt, LLAMA2_USE_POLICY.md, and NOTICE.