EmbodiWeave

Field notes / GraphHacks 2026 / Track 1

Try EmbodiWeave
in your browser.

Start training, watch the robot in 2D or 3D, and download the checkpoint and test results. The current 2D checkpoint passed validation and a fresh four-family confirmation. Target collisions were 8%, below the 10% limit. Its MuJoCo transfer gap and earlier failed attempts remain visible.

Follow the graph decision →Watch the 2-minute film
Live

Try the agent and graph

A fresh model chooses tools against running FalkorDB, proposes an adaptation and previews the gate. Exact calls and Cypher are visible.

Ask the live agent →
Recorded

PPO and Sonnet evidence

The replays, training outcomes and original Sonnet tool calls come from committed records. With a separate opt-in, the public agent can launch a new 60,000-step run.

Live

Train on the webpage

Choose 4,096, 60,000 or 150,000 steps. Watch the live simulator, then inspect 180 test episodes and download the checkpoint.

Run the full flow →

A judge’s route

Start with the measured decision.

Open the judge walkthrough for the earlier attempts, then inspect the newly confirmed checkpoint. Exact gates, live lineage queries and downloadable evidence are available. This recorded path does not depend on model capacity or a free trainer slot.

  1. 01

    See the original Sonnet replay on the same seed

    On seed 1300032, historical baseline P-static-b hits the pedestrian at step 51. The original Sonnet checkpoint reaches the goal at step 112. This is one paired episode, not the aggregate result.

    Replay in 3D or top view →
  2. 02

    Ask a fresh agent to investigate

    Choose “Explain the confirmed checkpoint” or write your own question. The model selects tools over MCP, queries live FalkorDB and proposes an origin. With separate training consent, it can launch a 60,000-step run after the guardrail checks the proposal. Open each receipt to see Cypher, parameters and returned evidence.

    Try the live agent → · Run Cypher directly
  3. 03

    Inspect the original Sonnet outcome

    The recorded agent used OpenRouter’s anthropic/claude-sonnet-5.5, 8 turns and 94 seconds. Target success rose from 37% to 87%, but collisions remained 13%, above the 10% limit. Three of four suites failed. The gate rejected promotion.

    Inspect the recorded tool calls →
  4. 04

    Change a premise without changing production

    In the studio, try a What-if. FalkorDB makes a GRAPH.COPY sandbox, inserts labelled hypothetical evidence, reruns selection and drops the copy. The response checks that production’s fingerprint is unchanged. Hypothetical decisions cannot launch training.

    Open the interactive studio →

Why FalkorDB is central

The graph is the agent’s working memory.

The planner depends on connected evidence, not just a list of models. If FalkorDB is unavailable, planning stops with a visible error.

Relationships answer the question

Which compatible starting point has measured transfer evidence for this mission? Which child did that trial produce, and did its complete, versioned validation clear the threshold?

The answer uses typed relationships, ancestry traversals and parameterized Cypher. The agent has read access through GRAPH.RO_QUERY; trusted tools append measured records.

Rebuildable, auditable evidence

Immutable JSON records are the recovery log. FalkorDB is the query and decision layer. Decisions store exact templates, parameters, returned record IDs and graph revision.

Final test results are kept outside the planning graph. Partial or custom evaluations cannot stand in for a complete canonical validation suite.

Explore the full schema and query →

Demo film v5 / a recorded action and its outcome

Follow the action.
Inspect what shipped.

The v2 agent trained 61,440 PPO steps and passed four validation gates across 180 episodes. Against its parent, target collisions rose from 2/60 to 3/60 and crossing from 0/40 to 1/40, so the child did not ship. V3 read that outcome but stopped before training. The closing success replay belongs to the shipped checkpoint, whose separate fresh confirmation had 8/100 target collisions against a 10% limit. The film is 118.10 seconds.

Narration: Deepgram Aura-2. Captions: Deepgram Nova-3. Download the SRT.

Robot motion is saved policy-driven 2D simulator output. Titles, graph layout and highlights are editorial. Browser captures replay completed evidence; waiting time is removed. The earlier studies remain below.

Download v5 in 1080p · Narration-free backup · v4 backup, unchanged · v3 backup · Timed script and sources

Music: “Cipher” by Kevin MacLeod (incompetech.com), CC BY 3.0. Edited excerpt and narration mix.

Earlier MuJoCo-v1 study / 40 episodes per suite

Earlier transfer failures stay visible.

MuJoCo target collisions were 24/40 for historical baseline P-static-b, 4/40 for original scratch and 3/40 for rehearsal. Rehearsal still collided in 8/40 crossing-only episodes. This is the earlier sim-to-sim transfer study, not a hardware safety claim. The latest study passed fresh 2D confirmation; its physics target still failed.

Query the live transfer graphRead the mapping and limitations

Reproduce locally

Real training. Your own next experiment.

Python 3.13, Docker, macOS or Linux. No GPU required. Try training on the live page. Use the local stack for larger budgets and custom experiments.

git clone https://github.com/himanshu748/embodiweave.git
cd embodiweave
docker compose up --build

Open http://127.0.0.1:8000/live.html. Docker starts the pinned CPU application and FalkorDB, with a persistent local evidence volume. Training works without a model API key. See the README for the unrestricted local development stack and optional model setup.

Checkpoint export bundles contain model weights, normalization, contract, records, Python source, source hashes, dependencies and a runnable evaluation command. New bundles use the repro-v3 suffix; historical bundles remain unchanged.

Full setup and verification →Download the original Sonnet checkpoint bundle

What this proves

Evidence, with its limits attached.

Demonstrated

  • A real LLM choosing MCP tools.
  • Real PPO checkpoints and measured simulator episodes.
  • FalkorDB evidence driving scratch versus parent selection.
  • A gate retaining the deployed policy after failed validation.
  • Equal budget comparisons, including a negative result.

Still preliminary

  • 2D policy training and a small MuJoCo transfer study, no physical robot.
  • Three training seeds per benchmark arm.
  • The earlier lineage failed held-out target collisions; the latest passed fresh 2D confirmation but still has a physics transfer gap.
  • No demonstrated efficiency or real world safety advantage.
  • Small checkpoint library; the selection rule was revised between experiments.