UCB

Octo-Base

UC Berkeley, USA · May 2024

Octo Model Team (UC Berkeley, Stanford University, Carnegie Mellon University, Google DeepMind)

Action policyOpen weights, MITAlso written Octo, Octo-Base 1.5, rail-berkeley/octo-base-1.5
Parameters
93 M
93M parameters (paper Table V: 12 layers, hidden size 768, MLP 3072, 12 heads).
GPU memory
0.2 GB
weights at bf16, computed. Fits an 8 GB card
Inference latency
not published
per action step
Weights
MIT
downloadable checkpoint

What it is

Octo-Base is a 93M-parameter transformer policy with a diffusion action head, pretrained on 800k Open X-Embodiment trajectories. It is not a VLA: language enters through a frozen t5-base encoder and there is no internet-pretrained vision-language backbone, which is why it generalizes far less than OpenVLA on hard out-of-distribution suites (20.0% versus 70.6% on the BridgeData V2 evaluation run by the OpenVLA authors). Its real value is adaptability and cost: about 100 demonstrations lift six new domains to a 72% average, new sensors and action spaces can be bolted on without touching pretrained weights, and the whole checkpoint is 0.81 GB. Known weak spots are wrist-camera processing and language conditioning, both traced to gaps in the pretraining mixture. The repository has been untouched since July 2024 and the paper is still at v2 from May 2024, so there is no newer version as of August 2026.

Architecture

Backbone
No VLM backbone. A transformer trained from scratch (12 layers, hidden size 768, MLP 3072, 12 heads) over shallow CNN patch encoders with 16 by 16 patches; language is encoded by a frozen pretrained t5-base (111M) text encoder.
Action head
Conditional diffusion head, a 3-layer MLP with hidden dimension 256, residual connections and layer normalization, trained with the DDPM objective on a cosine schedule with 20 diffusion steps. Only one transformer forward pass per action prediction; denoising happens inside the small head.
Action chunk
4 steps per forward pass
Control rate
13 Hz
Parameters
93M parameters (paper Table V: 12 layers, hidden size 768, MLP 3072, 12 heads). The sibling checkpoint Octo-Small has 27M. A 10M Octo-Tiny is mentioned in the scaling ablation (Section IV-C) but is not among the released checkpoints. The Hugging Face repo rail-berkeley/octo-base-1.5 holds a single 0.81 GB JAX checkpoint file. Version 1.5 is the current release and adds improved cross-attention and augmented language instructions over 1.0.
Pretraining data
800k robot trajectories from a hand-curated mixture of 25 Open X-Embodiment datasets. Pretraining took 8 hours on a TPUv4-128 pod for Octo-Small and 14 hours for Octo-Base, at batch size 2048 for 300K steps. Only 27% of the data contains wrist-camera images and only 56% carries language annotations, which the authors identify as the cause of two of the model's weaknesses.
Embodiments
WidowX (BridgeData V2), UR5, Google robot (RT-1 setup), Franka Emika Panda (Berkeley Peg Insertion, Stanford Coffee, CMU Baking, via fine-tuning), xArm (Berkeley Coke, via fine-tuning), ALOHA-style bimanual (Berkeley Bimanual, via fine-tuning)

What hardware it needs

This is the section most people came for, so it is worth being precise about which numbers are measured and which are arithmetic.

Computed from 93 M parameters. Weights only, so treat it as a floor: activations, image encoder intermediates and the CUDA context come on top, in practice 30 to 50 percent more for inference.
PrecisionWeightsNote
fp320.3 GBTraining master weights, rarely used for inference
bf16 or fp160.2 GBThe usual way these checkpoints are served
int80.1 GBQuantised, expect some loss of precision on fine motions
int40.0 GBAggressive quantisation, verify success rate before trusting it

Smallest real card that fits the bf16 weights plus a 40 percent runtime allowance: 8 GB. That is an estimate for inference, not for fine tuning. Full fine tuning also holds optimiser state and gradients, which typically costs several times the weights unless you use LoRA or a comparable adapter.

Published figures. These come from the model authors, not from this site.
Control rate13 Hz
Checkpoint download0.81 GB

Neither the paper nor the README states a GPU memory requirement. The abstract only claims Octo can be finetuned to new setups within a few hours on standard consumer GPUs. No exact number is published, so none is given here.

Reported results

Grouped by what the evaluation actually asked. Values from different suites are not comparable with each other, so each one keeps its suite and split.

Simulation

Reproducible benchmark suites. Everyone runs the same episodes, so the numbers can be compared inside a suite.

  • LIBERO average over Spatial, Object, Goal and Long after fine-tuning per suite, 3 seeds with 500 trials each
    75.1%
    success rateThird-party evaluation by the OpenVLA authors (Table 12), not by the Octo authors. Fine-tuned OpenVLA reaches 76.5% and Diffusion Policy from scratch 72.4%. The OpenVLA-OFT paper reuses this same 75.1% figure in its Table I.source

Real world

Rollouts on physical hardware. The setups differ, so read these as evidence, not as a ranking.

  • Octo real-robot finetuning suite (paper Table I) six new domains, about 100 target demonstrations each, 20 trials per domain, identical hyperparameters
    72%
    average success rateReported by the Octo authors. ResNet plus Transformer trained from scratch reaches 20% and a VC-1 pretrained visual representation 15% on the same setups.source
  • Octo real-robot finetuning suite (paper Table I) Berkeley Peg Insertion with an added force-torque proprioception input, 20 trials
    70%
    success rateReported by the Octo authors. From-scratch baseline 10%, VC-1 5%. This is the evidence that Octo can be finetuned onto a sensor modality absent from pretraining.source
  • Google robot (real RT-1 mobile manipulator) 12 tasks, 5 trials each, 60 rollouts
    26.7%
    mean success rateThird-party evaluation by the OpenVLA authors (Table 6), not by the Octo authors. OpenVLA reaches 85.0% and RT-2-X 78.3% on the identical tasks.source
  • BridgeData V2 (real WidowX) full 17-task evaluation suite, 170 rollouts
    20%
    mean success rateThird-party evaluation: measured by the OpenVLA authors (Table 4), not by the Octo authors, on a harder generalization suite than the Octo paper used. OpenVLA reaches 70.6%, RT-2-X 50.6%, RT-1-X 18.5%.source

Fine tuned tasks

No results in this category are published for this model.

On real hardware

Evaluated on 9 real robot setups across 4 institutions. Zero-shot on WidowX, UR5 and the RT-1 robot, where the paper reports a 29 point higher average success rate than RT-1-X and parity with the 55B RT-2-X on the tested WidowX and RT-1 tasks (Figure 5 is a bar chart, so exact per-robot values are not in the text). Fine-tuned results with about 100 demonstrations per domain: Berkeley Peg Insertion 70%, Stanford Coffee 75%, CMU Baking 50%, Berkeley Pick-Up 60%, Berkeley Coke 100%, Berkeley Bimanual 80%, averaging 72% over 20 trials each.

Fine tuning it yourself

JAX and Flax codebase. Fine-tuning updates the full model with the same diffusion objective as pretraining; the authors found this beats freezing subsets. Standard recipe: about 100 target demonstrations, 50k steps, cosine learning-rate decay with linear warmup, identical hyperparameters across all six evaluation domains. New observation encoders or action heads can be attached while keeping the pretrained transformer. Data must be converted to RLDS. The abstract claims fine-tuning takes a few hours on standard consumer GPUs but gives no memory figure.

Where it helps, where it does not

Strengths

  • At 93M parameters it runs on hardware where a 7B VLA does not, and the whole checkpoint is 0.81 GB
  • Flexible input and output spaces: it accepts language instructions or goal images, one or several cameras and optional proprioception, and new observation or action spaces can be attached at fine-tuning time without touching pretrained weights
  • Diffusion head models multi-modal action distributions; the authors show MSE heads produce hedging policies that move slowly and discrete heads lack grasp precision
  • Data-efficient adaptation: about 100 demonstrations and one shared hyperparameter recipe took six new domains from a 20% from-scratch baseline to 72%
  • Goal-image conditioning gives a 25 point higher success rate than language conditioning on WidowX, which is useful when tasks are hard to phrase

Limits

  • Struggles to use wrist-camera information; the authors report that fine-tuning was often stronger with a third-person camera alone than with both, and attribute this to only 27% of pretraining data having wrist images
  • Language-conditioned performance is markedly weaker than goal-conditioned performance because only 56% of the pretraining data is language annotated
  • Zero-shot performance degrades in new scenes and degrades heavily on novel behaviors such as flipping or precise insertion (Table VII analysis)
  • Trained and evaluated only on single and dual-arm manipulators; navigation and mobile manipulation are named as untried
  • In the harder OpenVLA generalization suites it collapses to 20.0% on BridgeData V2 and 26.7% on the Google robot, far behind internet-pretrained VLAs
  • No VRAM or latency figures are published beyond the 13 iterations per second on an RTX 4090, so hardware planning rests on that single data point
  • The repository has not been pushed since July 2024, so there is no maintenance activity to rely on

Sources

Everything on this page was taken from these documents. Where they disagree with what you read here, they win.

Entry last checked 2026-08-11.

Try a policy on a real arm

A physical SO-100 is online and free to drive in the browser, and five of the models in this arena can be fine tuned on a dataset you record with your own.