Octo

Select a versioned checkpoint

The source overview's quick-start example loads octo-base-1.5, while its checkpoint table links to unversioned Base and Small model names. The inference notebook uses Small 1.5. Record the exact model repository and revision used in an experiment rather than relying on the family name. The model card inspected here is for Base 1.5; its dimensions should not be assumed to describe a subsequently fine-tuned model. [S001] [S005] [S008]

Upstream's installation route uses Python 3.10 and JAX 0.4.20, with separate CUDA11 and TPU commands. Its initial debug fine-tuning command loads Small 1.5. Those pinned instructions need a compatible accelerator environment; this review did not install or execute them. The reported pretraining route is much larger: upstream describes approximately 1.2 TB of prepared data and TPU-pod training. An inference example does not reproduce that training process. [S001]

Match observations and history

Base 1.5's model card describes primary RGB images at 256 by 256 and wrist images at 128 by 128, with batch and history dimensions. It allows subsets of the listed observation and task keys and a history window of up to two timesteps. The environment adapter guide requires observation keys to match the keys used in the training configuration; absent keys are padded. [S006] [S008]

History padding is a different concern from a missing camera. The README explains that timestep_pad_mask marks which history timesteps exist, including the absent previous observation at the start of a trajectory. The per-modality mask handles missing observation elements. Using these mechanisms deliberately avoids presenting missing information as a real measurement. [S001]

The offline notebook adds batch and history dimensions explicitly, creates a language task through the model helper and samples an action chunk. It also passes action unnormalization statistics for the Bridge dataset. Those statistics belong to that example's data conventions; copying them into an unrelated robot integration does not establish the correct physical action interpretation. [S005]

An action chunk is not a complete plan

Base 1.5 predicts four seven-dimensional actions at a time. The README offers several execution choices: execute the chunk, execute only its first action before resampling, or use temporal ensembling. None of these turns a single output into a full task trajectory. Choose the action execution scheme together with the robot interface and evaluation configuration. [S001] [S008]

The fine-tuned ALOHA simulation example uses its own environment, checkpoint, normalization statistics and history/action wrappers. It also specifies separate simulator dependencies and an external ACT checkout. It is evidence of an integration example, not a universal adapter or proof of performance on another robot. No rollout, real-robot trial or inference-speed measurement was performed for this profile. [S007]

Model rights do not settle dataset rights

The source repository carries an MIT license, and the inspected Base1.5 model card separately reports MIT. The training mixture comprises multiple datasets, and the source links to their collection and preparation tooling. This review has not verified each constituent dataset's terms or downloaded the checkpoint and dataset payloads. It therefore makes no blanket licensing or complete reproduction claim for the training corpus. [S001] [S003] [S008]