Octo: An Open-Source Generalist Robot Policy
1. Paper overview
In one sentence: Octo makes a shared robot policy adaptable through modular token interfaces and diffusion action decoding, but its strongest transfer results still require target demonstrations and full-model finetuning. e01e02e03e04e05e08e10e12e16
| At a glance | What to know |
|---|---|
| Research problem | Source description Robot datasets differ in cameras, task descriptions, embodiments and action interfaces. A useful pretrained policy must transfer behavior while accommodating those differences. Octo targets low-level manipulation and inexpensive adaptation, with both language and goal-image task specification. e02e03 |
| Core mechanism | |
| A key reported result | Data-efficient finetuning across six robot setups: Reported average: Octo 72%. Per domain: Insertion 70%, Coffee 75%, Baking 50%, Pick-Up 60%, Coke 100%, Bimanual 80%. Task success rate; higher is better. About 100 target demonstrations per setup; common finetuning hyperparameters; 20 physical trials per domain. Average ResNet+Transformer Scratch: 20%; VC-1: 15%. The reported averages imply a 52-percentage-point advantage over scratch; the text calls this 52%. Octo’s and VC-1’s Average cells differ from the arithmetic means of the displayed domain cells (72.5% and about 15.8%, reader calculations). Reported cells are preserved. Architecture/objective differences prevent isolating pretraining alone. e08e09e10 |
| Reading caution | Author claim The authors report weak wrist-camera use and a language-versus-goal conditioning gap, attributing these partly to modality coverage: only 27% of training data has wrist cameras and 56% has language annotations. That explanation is not an isolated causal result. e14 |
Core contributions
- Author claim
The authors report releasing 27M-parameter Octo-Small and 93M-parameter Octo-Base checkpoints, JAX training/finetuning pipelines, and data loaders compatible with JAX and PyTorch. e15
Figure 2. A shared transformer accepts modular inputs and produces action readouts. Original paper, p. 3 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start at the left: language goes through a language encoder, while images become patches through a shallow CNN. The small p symbols mark learned positional embeddings. Follow the green task and blue observation blocks into the upper transformer; purple readouts feed the action heads. Section III-A supplies the attention rule that the schematic does not draw: observations see task tokens and current/past observations, while readouts can read those inputs without being attended to by them. In the lower panel, dashed blue and purple blocks identify interfaces added for a new sensor or action space. e03e05e07e08
What it supports. The reusable object is the transformer representation, with input and output interfaces adjusted around it. This explains how the same pretrained policy can initialize force-torque sensing or joint-control adaptations. The arrows terminate in actions; the diagram supplies no future-image predictor or model-based rollout used to choose those actions.
Where the evidence stops. The caption’s wording about adding interfaces without modifying pretrained parameters concerns weight reuse at initialization. Section III-C explicitly updates the full model during finetuning. The faded lower blocks therefore must not be read as a freeze mask.
2. Motivation
2.1 The problem and the proposed response
Robot datasets differ in cameras, task descriptions, embodiments and action interfaces. A useful pretrained policy must transfer behavior while accommodating those differences. Octo targets low-level manipulation and inexpensive adaptation, with both language and goal-image task specification. e02e03
2.2 What this reading follows
A generalist robot policy must cope with more than new objects: a new lab may change the camera arrangement, add force sensing, or command joints instead of an end effector. Octo addresses this through a common transformer representation with replaceable interfaces and an action diffusion head. Its evidence has three distinct layers: control in pretrained environments, adaptation to six new setups, and focused design ablations. Read the architecture first, then the training mixture, and keep those evaluation layers separate. This edition covers the supplied RSS 2024 proceedings PDF; its referenced appendices and extended diagnostic table are absent. e01e02e03e04e05e08e10e12e16
3. Research context
We place the paper in the collection through its world–action interface. The catalog labels and the reading’s assessment are shown separately.
| Catalog dimension | Recorded classification |
|---|---|
| Major category | VLA |
| Architecture | One Model |
| Prediction paradigm | Not applicable |
| Quadrant | Not applicable |
3.1 Evidence-based assessment
Supports the recorded classification
The shared transformer and integrated action readout support the catalog’s One Model policy interpretation despite modular tokenizers and heads. Language/image conditioning and diffusion action generation support VLA and the diffusion portion of its grouped subcategory; flow matching is not demonstrated. There is no future-observation rollout or inverse-dynamics action extraction, so the world-model prediction paradigm and quadrant remain not applicable. e03e05
This is the collection’s architectural analysis, not a new related-work survey. Benchmark comparisons and their protocols appear in Section 6.
4. Problem formulation
4.1 Inputs and outputs
| Inputs | Outputs |
|---|---|
|
|
4.2 Equations and their role
5. Method in detail
5.1 1. Make the interface modular without discarding the representation
Imagine moving from RGB-only control to Berkeley Insertion, where force-torque sensing becomes available. Octo’s design does not require concatenating everything into an immutable input format. Instead, task and observation streams become token blocks, and a new modality can obtain its own lightweight encoder and positional embeddings. The shared transformer remains a useful initialization. Its observation attention is causal across time, and learned readout tokens summarize preceding inputs without feeding information back into observation tokens. A changed action space can similarly receive a new output head. The distinction is between preserving an initialization and freezing a model: the actual adaptation recipe updates the full model. Reader interpretation: this modularity lowers the architectural cost of transfer, while target demonstrations still teach how the new interface should influence behavior. e03e07e08
Figure 3. Training exposure is a curated sampling distribution. Original paper, p. 4 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the title literally: these are dataset sampling weights, meaning the average contribution to a training batch. They are not a direct census of raw episodes. Locate the large Fractal, Kuka and Bridge sectors, then follow the leader lines to the smaller datasets. Section III-B explains the construction: select datasets with images and delta end-effector actions, exclude unsuitable or repetitive material, increase the weight of more diverse sources and downweight some repetitive ones. The plot belongs to a 25-dataset, 800k-episode training mixture, whose exact numerical weights are not printed on the sectors. e04e12e16
What it supports. The policy encounters a deliberately heterogeneous mixture rather than only demonstrations from the evaluation robot. That design is tested separately in Table II, where the full mixture performs better than the RT-X subset and Bridge-only alternative. The pie itself documents exposure, not the causal contribution of an individual dataset.
Where the evidence stops. No exact percentages should be inferred from sector areas. Detailed mixture information is deferred to Appendix C, absent from this PDF. The curation description also does not define a fully reproducible quantitative score for diversity.
5.2 2. Spend denoising computation on actions
After the transformer encodes the task and short observation history, Octo’s action head repeatedly transforms a Gaussian sample into a continuous action chunk. Equation (1) conditions each denoising step on the same readout embedding, the current sample and the denoising index. The expensive transformer therefore runs once per action prediction rather than once per denoising iteration. Training learns this conditional denoising behavior from noisy demonstration actions; it does not learn a future-image rollout used for planning. The authors argue that diffusion can represent multiple plausible action choices while preserving continuous precision. Table II is consistent with that explanation because diffusion exceeds MSE and discretized alternatives, but it does not separately measure multimodality or prove the explanation. The supplied PDF leaves the numerical denoising count and action-chunk horizon unresolved. e03e05e12e16
5.3 3. Separate useful transfer from unrestricted generalization
The three evaluation layers answer different questions. Figure 5 asks whether pretrained weights can execute language tasks in environments represented in the training mixture, with altered initial conditions. Table I asks whether those weights help after collecting about 100 demonstrations and adapting to a new setup, sometimes with new sensing or control interfaces. Table II then asks which recipe choices matter on a focused WidowX protocol. The strong finetuning averages should not be read as zero-shot performance on unseen robots. Likewise, the diffusion ablation should not be treated as a six-domain average. Reader interpretation: the strongest defensible conclusion is that Octo provides a useful, adaptable policy initialization. Broader claims require stronger held-out-task protocols, uncertainty reporting and the detailed evaluation material absent from the supplied proceedings artifact. e08e09e10e11e12e16
5.4 Training and inference
During training
The 800k demonstrations come from 25 Open X-Embodiment datasets. Curation excludes missing-image and non-delta-end-effector datasets, then filters repetitive, low-resolution or narrow-task data. More diverse datasets receive doubled weights; some repetitive datasets are downweighted. Missing camera channels are zero-padded, and gripper commands are aligned to 1=open and 0=closed. e04
Training adds Gaussian noise to demonstration actions and learns a conditional DDPM denoiser using a cosine noise schedule. Hindsight relabeling samples a future state uniformly as the goal image. Image augmentation and randomly dropping language or goal inputs support either task modality; demonstrations without language always use goal conditioning. e05e06
Octo-Base trains for 300k steps, batch size 2048, using AdamW with inverse-square-root learning-rate decay, weight decay 0.1 and gradient clipping 1.0. The reported run takes 14 hours on a TPU v4-128 pod. Finetuning uses full-model updates, 50k steps, cosine decay and linear warmup; a 24GB NVIDIA A5000 run takes approximately five hours. e06e07
During inference
Encode the observation history and task once. Starting from Gaussian noise, repeatedly denoise within the small action head while keeping the transformer readout embedding fixed. The resulting action chunk supplies robot commands. The supplied body does not specify the chunk length, denoising-step count, or how many actions execute before observations are refreshed. e03e05e16
5.5 Implementation flow
- Tokenize tasks and observations
A pretrained t5-base encoder supplies language embeddings; a shallow convolution stack converts images into flattened patches. Learned positional embeddings organize task tokens and successive observation blocks. Pretraining uses two observation frames. e03e06
- Read history without changing its representation
Observation tokens attend to task tokens and observations at the same or earlier time. Missing modalities are masked. Learned readout tokens attend to preceding task/observation tokens, but those tokens do not attend back to readouts. The readout embedding conditions the action head. e03
- Change interfaces, then adapt
New sensors or action spaces require new lightweight encoders, positional embeddings or output heads as appropriate. Existing transformer weights remain reusable at initialization. Figure 2’s wording about retaining pretrained parameters does not mean freezing them during optimization: Section III-C explicitly updates the full model. e03e07
6. Experiments & results
Octo learns a reusable visuomotor policy from 800k robot demonstrations. Modular input tokens feed a shared transformer, whose readout drives a diffusion action head. Its strongest evidence is adaptation across six physical manipulation setups with about 100 demonstrations each; flexibility still requires finetuning and does not imply universal zero-shot control.
The supplied proceedings PDF has all six main figures and Tables I–II, but references Appendices A–F and Table VII that it does not contain. Extended generalization diagnostics, complete configurations and detailed task success protocols therefore cannot be visually verified here. No claim is made about the contents or availability of an unsupplied supplement. e16
6.1 Read the original evidence
Figure 5. Zero-shot comparison within environments represented in pretraining. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Compare colors within each robot group, using the vertical success-rate axis and the parameter counts printed in the legend. Green Octo exceeds blue RT-1-X in all three groups. Peach RT-2-X appears only for WidowX and RT-1 Robot; an absent UR5 bar is not zero success. The hatched WidowX peach bar has special provenance: footnote 1 says it was taken from another study, while Octo and RT-1-X were tested on matching tasks. The text describes two language tasks per robot, ten trials per task, and varied initial conditions in pretraining-domain environments. e11
What it supports. The experiment supports useful direct robot control without target finetuning on these familiar setups. Octo is visually comparable to the much larger RT-2-X on the two available comparisons. This result concerns executing physical manipulation tasks, and should be kept distinct from the later experiments that adapt to new interfaces.
Where the evidence stops. The methods have different pretraining data and sizes: RT-X uses 350k robot episodes, versus Octo’s 800k. Exact bar values and uncertainty are not labeled. The text’s 29% aggregate wording is therefore not reinterpreted as a precise relative or percentage-point improvement.
Table I. Transfer gains span new sensory inputs, action interfaces and embodiments. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read down each task column before looking at the Average column. All three rows use the target domain demonstrations, but their starting representations and policy designs differ. Octo’s row is strongest in each domain. The retained protocol and footnotes matter: approximately 100 target demonstrations, 20 evaluation trials per domain, an asterisk for new force-torque observations, and daggers for joint-position action spaces. Berkeley Insertion therefore tests a sensory extension, while Pick-Up and Bimanual test changed action interfaces. The scratch baseline is a ResNet-based diffusion policy; VC-1 uses a pretrained visual encoder with an MSE action decoder. e07e08e09e10e16
What it supports. The reported Average cells are 72% for Octo, 20% for scratch and 15% for VC-1. Their first difference is 52 percentage points, not a 52% relative increase. Domain-level success supports the practical value of adapting the pretrained policy, including 70% on Insertion and 80% on Bimanual.
Where the evidence stops. Reader arithmetic gives 72.5% and about 15.8% from Octo’s and VC-1’s displayed domain cells, differing from the printed averages. Preserve the source cells. Architecture and objective differences also mean this table does not isolate pretraining alone; detailed protocols in Appendix F are absent.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| Data-efficient finetuning across six robot setups About 100 target demonstrations per setup; common finetuning hyperparameters; 20 physical trials per domain. | Reported average: Octo 72%. Per domain: Insertion 70%, Coffee 75%, Baking 50%, Pick-Up 60%, Coke 100%, Bimanual 80%. Task success rate; higher is better | Average ResNet+Transformer Scratch: 20%; VC-1: 15%. The reported averages imply a 52-percentage-point advantage over scratch; the text calls this 52%. Octo’s and VC-1’s Average cells differ from the arithmetic means of the displayed domain cells (72.5% and about 15.8%, reader calculations). Reported cells are preserved. Architecture/objective differences prevent isolating pretraining alone. e08e09e10 |
| WidowX design ablations Octo-Small; 40 trials across two language-conditioned and two goal-conditioned tasks. | Octo-Small: 83%. Aggregate task success rate | RT-X dataset mix: 60%; Bridge-only: 43%; discretized actions: 18%; MSE actions: 35%; ResNet-50 plus transformer: 70%. Diffusion decoding and the broad mixture perform best locally. Altering mixtures changes composition and scale; this does not establish a universal ranking across robots. e12 |
| Zero-shot manipulation on pretrained robot setups WidowX, UR5 and RT-1 Robot; two language tasks per robot and ten trials per task, with varying initial conditions. | Figure 5 shows Octo above RT-1-X on all three setups and similar to RT-2-X on the two tested setups. Physical task success rate | RT-X policies use 350k episodes versus Octo’s 800k. WidowX RT-2-X results come from another study using matching tasks. Zero-shot means no additional training in environments represented during pretraining. Bar heights lack exact labels; the text’s ambiguous 29% aggregate claim is not converted into a new estimate. e11 |
| Model-size scaling on UR5 and WidowX One language-conditioned task per robot; ten trials per task. | Figure 6 shows increasing success from Tiny (10M) through Small (27M) to Base (93M). Zero-shot task success rate | Three model sizes within the Octo family. A small real-robot scaling diagnostic without displayed uncertainty, rather than a fitted scaling law. e13 |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Table II. The action head and training mixture strongly affect this WidowX evaluation. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Use Octo-Small’s 83% as the reference row. The DATA group replaces the full mixture with the RT-X mixture or Bridge-only data. POLICY replaces diffusion decoding with discretized prediction or continuous MSE regression. ARCH replaces the transformer-first visual pathway with ResNet-50 plus a transformer. The original caption specifies 40 trials across two language-conditioned and two goal-conditioned WidowX tasks. This is a different protocol from both the language-only zero-shot comparison and six-domain finetuning. Section IV-C explains the intended comparisons, including the narrower 11-dataset RT-X mixture and the distinction between multimodal continuous actions and the alternative heads. e05e12
What it supports. The strongest local drops occur for discretized and MSE action prediction, which report 18% and 35% against Octo-Small’s 83%. The narrower data alternatives also underperform. These observations motivate examining the interaction between a varied demonstration mixture and an action distribution expressive enough to fit it.
Where the evidence stops. These are aggregate rates from one robot setup, without displayed uncertainty or per-task breakdowns. Data changes also alter scale and composition. The table supports the tested recipe but cannot by itself prove that multimodality is the sole cause of the diffusion advantage.
Figure 6. A small diagnostic supports increasing model capacity. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Follow each colored line separately as parameter count increases along the horizontal axis. Blue denotes UR5 and green denotes WidowX. The source text identifies Tiny, Small and Base as 10M, 27M and 93M parameters; the labels above the points mark those variants, not additional measurements. The caption narrows the protocol to one language-conditioned task per robot and ten trials per task. Read this plot as a check on model size within the Octo family. It is not the same task aggregate as Table II, even though both appear in the design-decision discussion. e13
What it supports. Both tested robots improve across the displayed size ladder. This supports the authors’ decision to offer a larger Base policy alongside Small, and suggests that capacity matters under their training recipe. It does not establish how success would scale on arbitrary tasks or with further increases in parameters.
Where the evidence stops. Only three sizes and a small number of physical trials are shown, without uncertainty or matched-compute analysis. The connecting lines are visual guides, not a fitted scaling law; the nearby model labels and leader marks are not error bars.
7. Analysis & limitations
7.1 What the evidence leaves open
The authors report weak wrist-camera use and a language-versus-goal conditioning gap, attributing these partly to modality coverage: only 27% of training data has wrist cameras and 56% has language annotations. That explanation is not an isolated causal result. e14
Evaluation covers single- and dual-arm manipulation; navigation, mobile manipulation and learning from suboptimal or online data remain future work. Small trial counts and absent uncertainty estimates limit reliability conclusions. e10e12e13e14
7.2 Questions for discussion
- Would diffusion still lead with matched action horizons, compute and target data? [e05, e12]
- Does improved wrist-camera coverage reverse the reported benefit of omitting that camera during finetuning? [e14]
8. Reproducibility audit
8.1 Requirements and known gaps
Reproduction needs matching data curation, camera/action preprocessing, target demonstrations and physical task protocols. Full mixture weights and hyperparameters are deferred to absent Appendices C–D, and evaluation details to absent Appendix F. Exact learning rates, action horizon, denoising count, software versions and detailed success criteria cannot be recovered from this PDF. e04e06e07e16
Proposed check: hold data, backbone, history and action horizon fixed while comparing diffusion, MSE and discretized heads on the four WidowX tasks, reporting per-task success and uncertainty across seeds. Test whether the Table II gap survives tighter controls. e05e12
8.2 Proposed reproduction checks
The following checks are proposals motivated by the paper. They have not been run as part of this reading.
Check 1: Does the diffusion advantage survive matched action interfaces?
Reader-proposed check, not performed: train Octo-Small variants with diffusion, MSE and discretized action heads using the same mixture, observation history, preprocessing, action horizon and backbone initialization. Use the Table II two-language/two-goal WidowX task design, with matched initial states and multiple training seeds; report each task separately with confidence intervals and inference latency. Document recovered hyperparameters rather than guessing the missing appendix settings. A stable diffusion advantage across tasks would support the claimed benefit of its action distribution; a gap that disappears under these controls would weaken that interpretation. e03e05e12e16
Check 2: Does the added force-torque input help after full-model adaptation?
Reader-proposed check, not performed: on Berkeley Insertion, hold target demonstrations, camera input, action head, training steps and evaluation starts fixed. Cross two factors: force-torque input present versus masked, and full-model finetuning versus training only the new interface/head. Use repeated seeds and the source’s 20-trial task format, reporting success and uncertainty for all four conditions. A force-input gain under full finetuning would support useful sensor transfer; a full-model gain over interface-only tuning would support adapting the shared representation. Similar scores would challenge those mechanisms even if overall success remained high. e03e07e08e10
8.3 Reading coverage
Visual audit: All 13 original PDF pages were rendered and visually inspected. Page 1 verifies title, credits, affiliations and Figure 1; page 2 the problem; pages 3–5 architecture, attention, data, Equation (1), optimizer and hardware details; page 6 setups, baselines, Figure 5 and its provenance footnote; page 7 both quantitative tables and protocols; page 8 scaling and limitations. Pages 9–13 complete the reference-only tail and confirm the absence of appendices in this artifact. All six final crops were individually viewed; narrow visuals were rendered at 300 DPI. Table I retains its original protocol caption because it contains the required input/action footnotes. Figure 2 arrows and added blocks were checked against Section III-A and the full-model update statement in III-C. Figure 5’s imported bar and Figure 6’s labels were checked against their captions. No numerical bar or sector values were reconstructed.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13. Appendix coverage: not present.
Original figures and tables remain the work of the source’s authors. Extractions preserve their scientific content; any HTML wrapper layout is disclosed with each figure. The surrounding reading notes are our own.
Text reading scope & known omissions
- Title, author block and abstract (p. 1)
- I. Introduction (pp. 1–2)
- II. Related Work (pp. 2–3)
- III. The Octo Model: A. Architecture, B. Training data, C. Training objective, D. Training Details, E. Model Checkpoints & Code (pp. 3–5)
- IV. Experiments: setups, comparisons, A. Out-of-the-Box Control, B. Data-Efficient Learning, C. Design Decisions (pp. 5–8)
- V. Discussion and Future Work (p. 8)
- References (pp. 8–13)
Outside the original text pass
- Identity notes: the title and author identities match the catalog. The inspected title page identifies RSS 2024, Delft, July 15–19, 2024, and credits Octo Model Team. It prints Homer Walke rather than Homer Rich Walke, Pannag Sanketi rather than Pannag R. Sanketi, and places Sanketi before Quan Vuong and Ted Xiao. Metadata preserves the observed credits. The catalog key octo_2023 does not establish a 2023 edition; no numbered revision or revision history is supplied.
- Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
- Separate supplemental material availability has not been fully verified.
- The extraction-only figure limitation was resolved by inspecting original PDF pages and all final crops. All six text chunks were read completely; all 13 PDF pages were visually inspected.
- The supplied proceedings PDF ends with references. Referenced Appendices A–F and Table VII are absent; contribution details, usage example, full mixture details, hyperparameters, extended ablations and detailed evaluation protocols were not reviewed.
- Code, checkpoints, external project pages and separate supplements were not inspected; no experiments were reproduced.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e01PDF p. 1, title/author/affiliation block and proceedings header
The exact title is Octo: An Open-Source Generalist Robot Policy. The RSS 2024 page credits Octo Model Team and 18 named authors, with UC Berkeley, Stanford, Carnegie Mellon University and Google Deepmind affiliations.
Go to primary source ↓e02PDF pp. 1–2, Abstract and Section I
Octo targets flexible generalist manipulation policies and adaptation across different sensors, tasks, robots and action spaces.
Go to primary source ↓e03PDF p. 3, Figure 2 and caption; p. 4, Section III-A
Modality tokenizers feed a block-masked transformer. Readouts passively attend to preceding inputs and feed action heads. Interfaces can be added while retaining pretrained transformer weights.
Go to primary source ↓e04PDF p. 3, data discussion; p. 4, Figure 3 and Section III-B; p. 5, opening paragraph
Octo curates 800k episodes from 25 datasets. Figure 3 depicts sampling weights; text describes filtering, diversity reweighting, camera padding and gripper alignment.
Go to primary source ↓e05PDF p. 5, Section III-C, Equation (1) and surrounding paragraphs
A single transformer pass conditions iterative action-space denoising. Gaussian initialization, conditional denoising and cosine scheduling define the DDPM policy.
Go to primary source ↓e06PDF p. 5, Section III-D
Training uses two-frame histories, future-state goal relabeling, augmentation and modality dropout. It reports optimizer settings, 300k steps, batch 2048 and 14 hours on a TPU v4-128 pod.
Go to primary source ↓e07PDF p. 4, Section III-A final paragraphs; p. 5, Sections III-C–D
Finetuning reuses transformer initialization, updates the full model for 50k steps, and uses cosine decay with warmup. Approximately five hours on a 24GB NVIDIA A5000 is reported.
Go to primary source ↓e08PDF p. 5, Section IV evaluation setups; p. 6, Figure 4 and caption
Nine physical setups span four institutions. Transfer includes force-torque input, joint-position actions, and new embodiments with about 100 in-domain demonstrations.
Go to primary source ↓e09PDF p. 6, Section IV comparisons
Scratch uses a ResNet/FiLM encoder and transformer diffusion decoder; VC-1 uses a pretrained ViT-B encoder and MLP trained with MSE. The scratch Octo transformer overfit small target datasets.
Go to primary source ↓e10PDF p. 7, Table I, all domain columns and Average column; caption and Section IV-B
Octo reports 70/75/50/60/100/80% across six domains, average 72%, versus baseline averages 20% and 15%. Each domain uses 20 trials; footnotes identify new inputs and actions.
Go to primary source ↓e11PDF p. 6, Figure 5, caption, footnote 1 and comparisons; p. 7, Section IV-A first paragraph
Figure 5 compares three robots, with RT-2-X on two. Text specifies tasks from pretraining domains, two tasks and ten trials each, 350k versus 800k episodes, and externally sourced WidowX RT-2-X results.
Go to primary source ↓e12PDF p. 7, Table II and caption, Section IV-C; p. 8, training-data and training-objective paragraphs
Octo-Small scores 83%; mix, policy and architecture alternatives score 60%, 43%, 18%, 35% and 70%. The caption specifies 40 WidowX trials across four tasks; the text explains compared mixtures and heads.
Go to primary source ↓e13PDF p. 8, Figure 6 and caption; Section IV-C model-scale paragraph
Tiny has 10M parameters, Small 27M and Base 93M. Figure 6 increases on both robots; the caption specifies one language task per robot and ten trials.
Go to primary source ↓e14PDF p. 8, Section V, limitations and future-work paragraphs
Wrist-camera and language coverage are 27% and 56%. The authors report modality weaknesses and discuss future learning beyond demonstrations and manipulation.
Go to primary source ↓e15PDF p. 5, Section III-E
The release statement lists Small/Base checkpoints, JAX training and finetuning scripts, and JAX/PyTorch-compatible data loaders.
Go to primary source ↓e16PDF p. 1, contribution footnote; pp. 3–7, references to Appendices B–F and Table VII; pp. 8–13, References through document end
The body refers to additional contribution, usage, mixture, hyperparameter, ablation and evaluation material. This proceedings artifact ends with references and contains none of those appendices or Table VII.
Go to primary source ↓8.5 Primary sources
Octo: An Open-Source Generalist Robot Policy ↗
PDF · 9,608 extracted words
Source fingerprint
2ad80d52e9ce79b6d3e052724631e4767a96a356850e48a5e5ef655a3b09a9c6