When Replanning Becomes the Bottleneck: Budgeted Embodied Agent Replanning

BRACE = Budgeted Replanning and Coordination for Embodied-Agents.

This work targets high-frequency replanning as a systems bottleneck for embodied agents: as context grows (history, perception summaries, retrieved memory), replanning latency develops a heavy tail and leads to deadline/SLO misses.

BRACE provides:

• A budgeted replanning controller (when to replan, what to include, and how to stay within a time/token budget), and
• Auditable phase logging (token + latency accounting on the replanning call path), and composes with efficiency modules (e.g., E-RECAP token pruning, optional retrieval/RAG).

Note: This repository accompanies our ICML 2026 paper.

Repo map

• brace/: BRACE controller core (budgeting + stability mechanisms)
• experiments/: domain runners (Habitat / RoboFactory / AirSim / proxy + stubs)
  • Includes benchmark helpers under experiments/robosuite/ and experiments/libero/
• analysis/: aggregation + audit tools (tables, schema coverage, trigger/controller audits)
  • Start here: analysis/README.md
• docs/: project page + guides
  • Website: docs/index.html
  • Main guide: docs/README.md
  • External benchmark notes: docs/EXTERNAL_BENCHMARKS.md
  • Controller spec: docs/CONTROLLER.md
  • Logging schema: docs/SCHEMA.md
  • Demo/media provenance: docs/PROVENANCE.md
  • E-RECAP guide: docs/e-recap.md
• configs/: curated configs
  • Defaults: configs/smoke/ (sanity checks)
  • Paper-facing: configs/experiments/ (curated eval/demo configs)
• scripts/: thin wrappers around Python entrypoints (smoke / run / postprocess)
• e-recap/: vendored E-RECAP module (optional)
• third_party/openmarl/: vendored OpenMARL components used by the VLA executor track

Quickstart (local smoke; no simulators required)

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

# Validates: run directory + schema + postprocess tables.
scripts/smoke_local.sh

Run (requires simulators / external assets)

These wrappers invoke the python entrypoints under experiments/ and write results to runs/:

# Meta AI Habitat (requires local habitat-setup + MP3D license assets)
scripts/run_habitat.sh --config configs/smoke/habitat_setup.json --run-name habitat_smoke

# RoboFactory (requires RoboFactory runtime + assets)
scripts/run_robofactory.sh --config configs/smoke/robofactory_lift_barrier.json --run-name robofactory_smoke

# Microsoft AirSim (requires UE binaries + BRACE_AIRSIM_ENVS_ROOT)
scripts/run_airsim.sh --config configs/smoke/airsim_multidrone_demo.json --run-name airsim_demo --ue-env airsimnh

Additional helper wrappers are also available for benchmark-native qualitative workflows:

# RoboSuite offscreen rendering / playback / demo collection
scripts/run_robosuite_render.sh --help
scripts/run_robosuite_playback.sh --help
scripts/run_robosuite_collect.sh --help

# LIBERO suite inspection / task rendering
scripts/run_libero_check.sh --help
scripts/run_libero_render.sh --help

Reproducibility & auditing (paper-facing tables)

This repo uses an on-disk, auditable run format:

• runs/<run_id>/run.json
• runs/<run_id>/events.jsonl
• runs/<run_id>/episode_metrics.jsonl

To postprocess a run into paper-facing tables:

scripts/postprocess_run.sh runs/<run_id>

This performs strict schema checks and writes markdown tables under artifacts/tables/. See:

• docs/SCHEMA.md for field definitions
• docs/PROVENANCE.md for committed demo media provenance

Citation

Coming soon.

Repo policy (open-source)

This repository is code + configs + docs only.

• Do not commit large weights/datasets/videos; keep them under your BRACE_MODELS_ROOT / BRACE_DATA_ROOT and reference via env vars.
• Local outputs are generated at runtime (e.g., runs/, artifacts/, data/) and are not shipped in this repo. See docs/LOCAL_OUTPUTS.md.