Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 83 of the 95 references it cites were found in a registry or an index; 11 of the 31 sources its claims cite were checked for support.

Design and evaluation of a global workspace agent embodied in a realistic multimodal environment

Open access
Dossa, Rousslan Fernand Julien, Arulkumaran, Kai, Juliani, Arthur, Sasai, Shuntaro, Kanai, Ryota · 2024 · Frontiers in Computational Neuroscience

Abstract

As the apparent intelligence of artificial neural networks (ANNs) advances, they are increasingly likened to the functional networks and information processing capabilities of the human brain. Such comparisons have typically focused on particular modalities, such as vision or language. The next frontier is to use the latest advances in ANNs to design and investigate scalable models of higher-level cognitive processes, such as conscious information access, which have historically lacked concrete and specific hypotheses for scientific evaluation. In this work, we propose and then empirically assess an embodied agent with a structure based on global workspace theory (GWT) as specified in the recently proposed “indicator properties” of consciousness. In contrast to prior works on GWT which utilized single modalities, our agent is trained to navigate 3D environments based on realistic audiovisual inputs. We find that the global workspace architecture performs better and more robustly at smaller working memory sizes, as compared to a standard recurrent architecture. Beyond performance, we perform a series of analyses on the learned representations of our architecture and share findings that point to task complexity and regularization being essential for feature learning and the development of meaningful attentional patterns within the workspace.

Overview

CoChat Summary

TL;DR

The authors design an embodied global-workspace agent satisfying GWT indicator properties and train it on realistic audiovisual 3D navigation. It outperforms a GRU baseline mainly at smaller working-memory sizes, while larger workspaces yield little benefit and attention patterns become less dynamic.

Key findings

  • The global workspace agent achieved ~80% success across all working-memory sizes apart from size 256, where some runs performed poorly and were considered possible outliers; the baseline GRU achieved only over 60% across all memory sizes and became competitive only at higher sizes. For reference, a random policy achieved 0% and the original SAVi RL agent reached 25% on average.
  • Ablating working-memory size from 32 to 512 showed negligible performance differences for the global workspace agent, suggesting the task’s short-term memory demand is small and that SAVi is not an adequate benchmark for working-memory capacity; larger agents did not perform significantly better.
  • Attention analyses showed all agents primarily used cross-attention across input modalities; smaller, more bottlenecked agents developed more mixed attention patterns integrating modalities over time, whereas larger workspaces produced more binary attention that saturated early, with working memory and null keys largely unattended.
  • Probing found acoustic features encoded target object category and visual features encoded scene; target category from visual features ranged 14–37% versus chance 4%, scene from audio ranged 6–14% versus chance 1.79%, and after acoustic cue cutoff target-category probing accuracy dropped abruptly for both variants, though for memory sizes ≤128 global workspace agents had higher average probing accuracy than GRU counterparts. Broadcast weight analysis indicated previous working memory was prioritized over current sensory inputs in both audio and vision encoders.

Methods

  • Architecture: an embodied agent with independent visual and acoustic CNN+GRU encoders, a recurrent working memory/global workspace, cross-attention bottleneck with a null input, and global broadcast from previous working memory to encoders; the baseline GRU ablates global broadcast and cross-attention while retaining independent encoders and recurrence.
  • Task and data: SoundSpaces/SAVi semantic audiovisual navigation in 3D scanned real environments, with 21 target object categories and 56 training scenes; observations were 128×128×3 RGB images and 65×25×2 spectrograms, with four discrete actions (forward, turn left, turn right, stop).
  • Training and analysis: behavioral cloning on 500,000 oracle tuples, batch size 10 contiguous trajectories of length 150, entropy coefficient 0.2, Bayesian hyperparameter optimization, final five seeds for 50,000 iterations across memory sizes 32, 64, 128, 256, and 512; each run took ~5 days on an NVIDIA RTX 3090, requiring more than 8,616 GPU hours total. Evaluation used interquartile mean ±95% CI over the last 50 evaluations, probing MLPs, attention weights, and normalized input-weight magnitudes.

Limitations

  • Stated: SAVi is not an adequate benchmark for testing working-memory capacity; a more sophisticated task or environment is needed; developing realistic simulator environments is labor-intensive and beyond the scope of this work. Performance sometimes declined after the halfway point, which may be an artifact of hyperparameter optimization runs stopping early to reduce computational requirements by ~4,000 GPU hours.
  • Stated: neither the conventional GRU nor the proposed attention mechanism preserved high enough target-category probing accuracy after sound cutoff, contrasting with the hypothesized ability of the global workspace to sustain a representation over time; no consistent attention patterns appeared across runs, and all agents achieved similar task performance despite different attention.
  • Evident: findings come from one audiovisual navigation task using imitation learning from an oracle rather than online RL; the work evaluates indicator properties, not consciousness itself; and the baseline retains some GWT/RPT properties (GWT-1, RPT-1/2), so the ablation is partial.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.3389/fncom.2024.1352685

Cites

83 of 89 references verified6 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Figure 1 Diagram of agent architectures. (A) Baseline GRU agent with recurrent encoders. (B) Global-workspace agent satisfying all RPT and GWT indicator properties. The double bars marking the path from the past step's working memory in the recurrent encoder for each modality emphasize the gradient-stopping operation during training that was required to stabilize the learning of this agent variant. (C) The proposed cross-attention mechanism takes as input the past step's working memory, the current visual and acoustic features to be used as query sources, and then computes an attention mask for each of those items, along with a null input; this therefore simultaneously incorporates bottom-up, cross, self, and top-down attention. Intuitively, the attention mask determines which information to incorporate from each input.
Figure 1 Diagram of agent architectures. (A) Baseline GRU agent with recurrent encoders. (B) Global-workspace agent satisfying all RPT and GWT indicator properties. The double bars marking the path from the past step's working memory in the recurrent encoder for each modality emphasize the gradient-stopping operation during training that was required to stabilize the learning of this agent variant. (C) The proposed cross-attention mechanism takes as input the past step's working memory, the current visual and acoustic features to be used as query sources, and then computes an attention mask for each of those items, along with a null input; this therefore simultaneously incorporates bottom-up, cross, self, and top-down attention. Intuitively, the attention mask determines which information to incorporate from each input.CC BY · PubMed Central
Figure 2 (A) Audiovisual navigation task provided by SoundSpaces. The agent has to navigate the rooms to reach the source of the noise (a ringing phone), using a single RGB camera view and binaural audio spectrograms as input. Figure reproduced from Chen et al. (2020) with the authors' permission. (B) Two example trajectories of the SAVi (SoundSpaces extension) navigation task. In SAVi tasks, the source of noise can originate from target objects of 21 different categories. In the first trajectory (magenta), the agent hears the sound of a dripping sink, while in the second trajectory (orange), it hears the sound of a cabinet door that is either opened or closed. The duration of the acoustic cues varies. Figure reproduced from Chen et al. (2021) with the authors' permission.
Figure 2 (A) Audiovisual navigation task provided by SoundSpaces. The agent has to navigate the rooms to reach the source of the noise (a ringing phone), using a single RGB camera view and binaural audio spectrograms as input. Figure reproduced from Chen et al. (2020) with the authors' permission. (B) Two example trajectories of the SAVi (SoundSpaces extension) navigation task. In SAVi tasks, the source of noise can originate from target objects of 21 different categories. In the first trajectory (magenta), the agent hears the sound of a dripping sink, while in the second trajectory (orange), it hears the sound of a cabinet door that is either opened or closed. The duration of the acoustic cues varies. Figure reproduced from Chen et al. (2021) with the authors' permission.CC BY · PubMed Central
Figure 3 Success rate of each agent type, across a range of working memory sizes. We report the IQM ±95% CI of the rolling average over the last 50 evaluations, aggregated over five independent runs (random seeds). The performance of the global workspace agent is more robust, particularly for smaller working memory sizes.
Figure 3 Success rate of each agent type, across a range of working memory sizes. We report the IQM ±95% CI of the rolling average over the last 50 evaluations, aggregated over five independent runs (random seeds). The performance of the global workspace agent is more robust, particularly for smaller working memory sizes.CC BY · PubMed Central
Figure 4 Probing accuracy for the target object category and scene, conditioned on the working memory, visual features, and acoustic features respectively, grouped by working memory sizes and agent variants. Accuracies are averaged over all evaluation trajectories and five independent runs, with error bars representing ±1 standard deviation.
Figure 4 Probing accuracy for the target object category and scene, conditioned on the working memory, visual features, and acoustic features respectively, grouped by working memory sizes and agent variants. Accuracies are averaged over all evaluation trajectories and five independent runs, with error bars representing ±1 standard deviation.CC BY · PubMed Central
Figure 5 Probing accuracy for the target object category based on the agent's working memory, across five different trajectories. The gray area delineates the part of each trajectory where the acoustic cues are playing.
Figure 5 Probing accuracy for the target object category based on the agent's working memory, across five different trajectories. The gray area delineates the part of each trajectory where the acoustic cues are playing.CC BY · PubMed Central
Figure 6 Probing accuracy for an episode's scene based on the agent's working memory, across five different trajectories. The gray area delineates the part of each trajectory where the acoustic cues are playing.
Figure 6 Probing accuracy for an episode's scene based on the agent's working memory, across five different trajectories. The gray area delineates the part of each trajectory where the acoustic cues are playing.CC BY · PubMed Central
Figure 7 Overview of the attentional patterns of the global workspace agents, based on the working memory query, over five unrolled trajectories.
Figure 7 Overview of the attentional patterns of the global workspace agents, based on the working memory query, over five unrolled trajectories.CC BY · PubMed Central
Figure 8 Attention weights averaged over all 150 evaluation trajectories and independent runs (seeds) for each query and key-value combination, grouped by working memory size.
Figure 8 Attention weights averaged over all 150 evaluation trajectories and independent runs (seeds) for each query and key-value combination, grouped by working memory size.CC BY · PubMed Central
Figure 9 Normalized average magnitude of the input weights for the audio and vision recurrent encoders, for the sensory inputs vs. the previous working memory, grouped by working memory size.
Figure 9 Normalized average magnitude of the input weights for the audio and vision recurrent encoders, for the sensory inputs vs. the previous working memory, grouped by working memory size.CC BY · PubMed Central

Verification

Claims the paper makes

7 Valid · 4 Not valid · 12 Not checked · 63 Own results and interpretations

37 more claims · 1 with their cited sources checked

Create a free account to see all 86 claims and to verify the 3 references still open.

See all claims — free account
  • It may state two separate claims as one. 26
  • Probably not a substantial claim. 23

Create a free account to see these claims.

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

2 of 2

Citations

Cites

Show all 95 references

Research works printed but not linked: 3

  1. Baker, 2019, Emergent Tool use from Multi-Agent Interaction
  2. Goodfellow, 2016, Deep Learning
  3. 2018, OpenAI Five

Cited by

See all 5

Metadata

Authors
Dossa, R. F. J., Arulkumaran, K., Juliani, A., Sasai, S., & Kanai, R.
Affiliations
  • Araya Inc. (Japan) (JP)
  • Microsoft Research New York City (United States) (US)
Corresponding author
Dossa, Rousslan Fernand Julien
Venue
Frontiers in Computational Neuroscience
Published
2024-06-14
Volume
18
Pages
1352685-1352685
ISSN
1662-5188 (ISSN-L)
OpenAlex
W4399687862
Cited by
5
References
95
Citations by year
2026: 2 · 2025: 3
FWCI
0.73
Type
journal article
Language
en
Open access
Locations
  • Frontiers in Computational Neuroscience (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Psychiatry, Mental Health, Neuroscience
  • Social Sciences › Arts and Humanities › Philosophy › Philosophy and Theoretical Science
  • Life Sciences › Neuroscience › Cognitive Neuroscience › Neural dynamics and brain function
Keywords
global workspace theory, consciousness, artificial neural networks, working memory, multimodal environments, recurrent neural networks, attention, representation learning, regularization, task complexity

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-13, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-13 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol and qwen3.8-omni-flash).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Design and evaluation of a global workspace agent embodied in a realistic multimodal environment. https://app.cochat.ai/papers/10.3389/fncom.2024.1352685
The paper
Dossa, R. F. J., Arulkumaran, K., Juliani, A., Sasai, S., & Kanai, R. (2024). Design and evaluation of a global workspace agent embodied in a realistic multimodal environment. Frontiers in Computational Neuroscience. https://doi.org/10.3389/fncom.2024.1352685

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.