AMD University Program Datawhale-EAIPhysical AI Evidence Lab
AMD MIGRATION CASEBOOK

Turn migration work into reusable capability

Datawhale-EAI connects runtime, simulation, data, policy, rendering, and evaluation into a repeatable AMD Radeon and ROCm migration method. Every case links its source, media, result files, and reproduction entry point.

Project overview Migration playbook Engineering blog
01 / METHOD

A six-layer migration method

Each layer exposes a clear contract, turning one successful port into the starting point for the next robot stack.

01

Runtime

Pin ROCm, PyTorch or JAX, MuJoCo, simulator, and plugin versions into a stable AMD environment.

02

Simulation

Run official scene reset, step, render, and video export while preserving task geometry and timing.

03

Data

Align trajectory dimensions, temporal order, episode boundaries, statistics, and training schemas.

04

Policy

Connect official weights, training logs, checkpoints, and resume semantics to the native workflow.

05

Replay

Unify observation/action bridges, cameras, control timing, and headless rendering into readable media.

06

Evidence

Archive fixed tasks, seeds, JSON, SHA-256, and representative videos as a downloadable result package.

02 / SIMULATOR

DISCOVERSE: simulation, expert data, and multi-view replay

18 / 18 runtime gates

DISCOVERSE forms the broadest simulator workstream in the suite: MuJoCo tasks, AIRBOT and MMK2 robots, expert trajectories, policy entry points, ROS2, LiDAR, 3DGS, and MP4 export share one AMD engineering path.

  • Runtime18/18 core gates; AIRBOT 12/12; MMK2 8/8
  • Expert trajectoriesblock_bridge_place strict replay 31/31; 500 expert episodes
  • 3DGSROCm renderer loads 1,334,537 RM2 points and 1,026,855 SkyRover points, plus dynamic Franka and UR5e replay
  • Media and dataOpen the 3DGS gallery
  • Hardware interfacesRealSense, gamepad, robot-arm, and ROS2 interfaces retain the upstream contracts
500expert episodes31 / 31strict replay frames23DGS targets
03 / JAX RUNTIME

ROCm JAX and OpenPI: native policy execution

JAX 0.10.0 verified

Pi0 and Pi0.5 use JAX and Flax in OpenPI. The AMD-native route connects GPU kernels, Orbax restore, tokenizer, normalization statistics, model inference, headless MuJoCo, and fixed-task evaluation.

  • Official baseAMD ROCm JAX installation guide
  • RuntimePython 3.12, ROCm 7.14, JAX/JAXlib 0.10.0, rocm:0 on gfx1151
  • Model pathOfficial Pi0.5 checkpoints, tokenizer, normalization, vision inputs, and action decoding
  • EvaluationRoboCasa365 16-task panel and DexJoCo 11-task multi-task archive
0.10.0ROCm JAX16 × 50RoboCasa archive11DexJoCo tasks

One evidence contract across JAX and PyTorch

Native JAX policies and ROCm PyTorch workflows share checkpoint identity, task protocol, video, runtime manifests, and SHA-256 validation.

04 / BENCHMARK

RoboCasa365: one protocol for household manipulation

16 official tasks

The AMD 395 runs the official RoboCasa365 assets, environments, Pi0.5 and GR00T policies, MP4 export, and a shared 16-task evaluation. Synchronized center, left, right, and wrist cameras capture long-horizon behavior.

  • Protocol16 official tasks, 50 episodes per task and model
  • PoliciesGR00T N1.5 official 120k and Pi0.5 official 75k checkpoints
  • Four-view replay1920x1080 at 20 fps for task-level review
  • ArtifactsPer-task stats, aggregate JSON, videos, manifests, and hashes
16official tasks50episodes / task4camera views
05 / POLICY STACK

From model training to closed-loop evidence

The policy workflows share notebooks, checkpoint audits, evaluation exports, and result indexes for repeatable AMD runs.

Every Embodied

SmolVLA, Pi0, and ACT notebooks cover ordinary training, protected training, evaluation, and video generation.

Notebook archive

RoboWits

The W7900 ACT 100k route preserves checkpoints, HD replay, evaluation data, and SHA records.

100k route

Public package

Source, report, website, result JSON, representative video, Hugging Face weights, and hashes form one review path.

Evidence ready
06 / SAFETY CONTROL

PAC-MAN: whole-body dodging with the upstream G1 policy

960-D → 29 joints

The reproduction connects four-frame projectile depth, G1 proprioception, the upstream dodge_link_cbf ONNX actor, 29 joint targets, balance control, and free-base dynamics. One policy produces a duck and left/right sidesteps from a standing command.

  • Observation384-D proprioceptive history plus 576-D projectile-depth history
  • Motion18.8 cm duck, 25.5 cm right sidestep, 31.0 cm left sidestep
  • Output1080p film, motion metrics, runtime manifest, and SHA-256
Open the PAC-MAN case study
3 / 3dodge modes13 swhole-body film1080pfree-base replay
07 / DEXTEROUS POLICY

DexJoCo: multi-task dexterity on native ROCm JAX

11-task archive

Pi0.5 connects official weights to eleven dexterous tasks

The native JAX 0.10 path restores official Orbax parameters and runs bimanual coordination, tool use, fine grasping, and long-contact tasks. The archive links the formal seed-0 panel, deterministic first-success traces, and task videos.

brave-eai/dexjoco · DexJoCo/DexJoCo-Pi05