AMD Radeon Hackathon 2026 · Track 3: Physical AI Challenge · Liu Liang
Warehouse AMR simulation on AMD Radeon and ROCm
Watch the 3:59 demo without downloading · Explore the live Evidence Console · Read the technical report · Reproduce the evidence
The full video is also stored in this repository,
with a hash-verified Release mirror.
The player, results pages, data adapters, tests, and complete website source are
all in showcase/.
This is a silent excerpt of the recorded simulation replay—not a new experiment or physical-robot footage. Click the moving preview to jump to the complete 3:59 demo on the project site, or play the 13-second 1080p MP4 directly.
A loaded warehouse carrier is approaching a corridor. Its first RGB-D view is ambiguous. Six software Claims may appear to support the route, but if they all come from the same camera capture, they are still only one physical observation.
Look Twice checks whether the evidence is good enough for the next physical action. If it is not, the system explains what is missing, sends a low-risk scout to collect a useful new view, and evaluates the same action again. The carrier moves directly only after the evidence passes both authorization paths; otherwise it detours or remains denied.
AMD-accelerated Genesis RGB-D
↓
lineage-aware Claims + calibrated prediction set
↓
Action Contract for this robot, payload, corridor and time window
↓
Python policy ∧ Purify Go contract gate
↓
direct motion, targeted evidence repair, or safe detour
The main loop is simple: deny → diagnose → acquire new evidence → re-qualify.
RGB, depth, copied artifacts, and forwarded Claims from one capture cannot vote as independent observations. Look Twice traces them back to physical capture roots before counting support.
A model score is not permission to move. Evidence must be fresh, calibrated, in scope, and sufficiently independent for a named robot, payload, corridor, action, and validity window.
A denial includes a machine-readable reason: missing viewpoint, weak independence, stale evidence, unresolved prediction set, or another bounded gap. The scout uses that gap to choose the next observation instead of blindly waiting or collecting more of the same view.
Split-conformal prediction sets carry uncertainty into the gate. An ambiguous set or an out-of-scope calibration cannot silently become a clear corridor.
Purify is the standalone Go reference contract core in
purify_robotics/, not another perception model. Direct
motion requires agreement between the Python policy and this separately
compiled gate. A missing service, failed clause, or disagreement keeps the
action closed.
Each GateReceipt records the Claims used and discounted, physical roots,
clause results, BeliefGaps, validity window, and content hash. Later evidence
can invalidate an earlier plan without erasing its history.
The headline comparison uses paired simulated warehouse worlds with the same initial condition for active and passive policies.
| Evidence tier | Result |
|---|---|
| Permanent locked V8 test | Active direct route 11/12; passive 0/12. Both completed 12/12 missions. |
| Publicly preregistered 30-world challenge | Active full-chain direct 29/30; passive 0/30. 60/60 missions completed with 0 recorded unsafe episodes and 0 fallbacks. The one dual-blocked world took the safe detour. |
| Active-view efficiency supplement | Both selectors preserved 20/20 direct outcomes. The candidate reduced mean scout path 40.3%, mean team path 15.0%, and physical captures 27.5% relative to the baseline selector. |
| Wheel-dynamics replay | 30/30 archived decisions passed across 30 serial three-robot scenes: 30 scouts, 30 active carriers, and 30 passive carriers reached their goals using wheel-speed control after build. |
| Solver-scale complement | The same 30 fixed decisions passed as 20/20 in one 60-body scene and 10/10 in one 30-body scene: 90 cumulative distinct robots, with at most 60 co-resident. |
The preregistered challenge is the primary public supplement. The efficiency, dynamics, and scale rows are separate same-generator simulation studies; they do not replace the locked result or claim physical-robot validation.
Evidence:
- Compact challenge evidence card
- 30-world challenge report
- Active-view efficiency result
- 30-seed decision-bound dynamics result
- 60-body prefix report
- 30-body suffix report
- Same-host CPU/GPU comparison result
- Raw challenge archive
| Layer | Role in Look Twice |
|---|---|
AMD Radeon gfx1100, 48 GiB |
Competition GPU environment |
Genesis 1.1.2 with gs.amdgpu |
Warehouse simulation and RGB-D rendering |
| PyTorch 2.9.1 ROCm / HIP 7.2 | Spatial RGB-D inference and tensor processing |
rocm-smi telemetry |
Full-run utilization, memory, power, and temperature record |
| Purify Go core on CPU | Deterministic contract evaluation and receipt generation |
All 60 preregistered challenge episodes loaded the frozen checkpoint, rendered live Genesis RGB-D, ran spatial inference through the ROCm stack, and emitted Go receipts. The complete 1,685.5-second wall is covered by 844 retained telemetry samples, including idle time. See the AMD environment, telemetry, and technical report for the exact configuration and measurement boundaries.
In a fixed additive comparison on that same host, the frozen FP32 model ran 1.76–2.04× faster on Radeon than the 64-thread CPU path across batch sizes 1, 4, and 8. Each point used 20 warm-ups and 100 synchronized measurements; all output tensors matched at the declared tolerance. This is model-forward performance, not end-to-end robot latency. See the result and boundary, precommitted protocol, and formal JSON.
The public site is a browser-native replay of recorded evidence. It does not need a GPU or a private service:
cd showcase
npm ci
npm test
npm run dev
# open http://localhost:3000Run the core evidence checks on CPU:
python3 scripts/build_competition_replays.py
python3 scripts/build_competition_hook.py
python3 -m unittest tests.test_competition_replay -v
cd purify_robotics
go test ./...The original source-SHA guard is tied to the immutable frozen-foundation tag, not to this later additive branch:
git clone --branch v8-competition-final-2026-08-05 --single-branch \
https://github.com/eason4kim-rocket/look-twice.git look-twice-frozen
cd look-twice-frozen
python3 scripts/verify_frozen_foundation.pyOr launch the packaged console from the repository root:
docker compose up --build
# open http://localhost:3000The full reproduction guide covers the challenge archive, locked-input verifier, dynamics reports, scale reports, AMD runtime, and frozen checkpoint identity.
The repository includes the public schemas, Go contract core, Python policy and repair planner, validators, deterministic replay builder, evidence data, and the full website.
During this work, a Genesis URDF inertial-origin issue was isolated and reported upstream as issue #3183. The focused fix is open as PR #3184, with its bounded validation record. The PR is open and unmerged; no upstream acceptance is claimed.
| Path | Contents |
|---|---|
showcase/ |
Complete bilingual evidence website, native video player, data, and tests |
src/ |
Robot loop, RGB-D model, Claims, conformal qualification, and repair planning |
purify_robotics/ |
Standalone Go Action Contract reference core |
schemas/ |
Public Claim, Action Contract, receipt, and BeliefGap schemas |
release/v8-frozen/ |
Frozen V8 manifests, reports, and guarded evidence import |
release/v8-derived/ |
Clearly separated additive simulation evidence |
scripts/ |
Reproduction, verification, replay, benchmark, and packaging tools |
docs/ |
Technical report, architecture, evidence boundaries, and validation notes |
submission/official-repo/ |
Self-contained Track 3 submission package prepared for review |
Look Twice V8 is a simulation research prototype for warehouse AMRs. Its
formal result is the frozen candidate; later supplements are labeled
separately. In the primary kinematic runs, carrier and scout are logical poses,
viewpoints, and capture roots on one shared Genesis chassis—not two robots
moving simultaneously. The separately labeled wheel-dynamics studies
instantiate non-fixed bodies. The project does not claim real-robot
deployment, sim-to-real or OOD performance, simultaneous 90-robot fleet
control, formal safety proof, or certification. Content hashes detect changes
but do not prove that a sensor's original observation was truthful. Full boundaries are documented in
docs/V8_EVIDENCE_BOUNDARY.md.
Built by Liu Liang for Track 3 of the AMD Radeon Hackathon 2026.
Apache-2.0. NOTICE defines the public Purify reference-core boundary.