SYNAPSE-Bench

SYNAPSE-Bench is a multimodal benchmark for evaluating cooperative planning, action selection, and communication in embodied multi-agent tasks. Each example describes one agent's local decision context in a household-like environment, including visual observations, structured belief state, teammate model, coordination dependencies, interaction history, a local scene graph, and the reference next-step response.

The released data is formatted for vision-language and language-agent evaluation. Models are expected to choose both an embodied action and a communication act that keeps the teammate coordinated.

Repository Layout

SYNAPSE-Bench_release/
├── README.md
├── synapsebench100.json
├── videos.zip
└── code/
    ├── README.md
    ├── get_interactive_rollouts.py
    ├── get_interactive_rollouts_s.py
    ├── graphstitch.py
    ├── eval_interactive_rollout.py
    ├── eval_patterns.py
    └── exp_interactive_rolllout.sh

Data Files

• synapsebench100.json: 100 evaluation sequences, stored as 834 turn-level records. The records form 417 paired decision steps: one row for humanoid_01 and one row for quadruped_01 at each step.
• videos.zip: scene-level visual and metadata assets. The directory contains scenes with raw camera videos, instance segmentation metadata, and local relation files.
• code/: rollout generation and evaluation scripts.

Benchmark Statistics

For synapsebench100.json:

Quantity	Value
Sequences	100
Turn-level records	834
Paired decision steps	417
Agents	humanoid_01, quadruped_01
Records per agent	417 each
Unique scenes referenced	105
Visual references per record	4
Sequence length range	3-10 paired steps

Task Format

Each top-level JSON record has this structure:

{
  "id": "0.000000f9725d40",
  "images": [
    "<sampled_frames_from_agent's_pov>"
  ],
  "conversations": [
    {
      "from": "human",
      "value": "{... JSON-encoded agent state ...}"
    },
    {
      "from": "gpt",
      "value": "{... JSON-encoded reference response ...}"
    }
  ],
  "done": false,
  "local_scene_graph": [
    ["humanoid_01", "medical_kit", "close_to"]
  ]
}

Top-Level Fields

• id: unique record identifier.
• images: relative visual observation references for the current agent and timestep.
• conversations: two-message instruction/answer pair. The human value is the model input, and the gpt value is the reference output.
• done: marks the terminal paired step of a sequence. Both agent rows in the final pair have done=true.
• local_scene_graph: local relational graph as triples of [subject, object, relation].