Datawhale-EAI · AMD Radeon + ROCm

AMD Physical AI for embodied household manipulation

A complete workflow from data and training to closed-loop evaluation, anchored by the RoboCasa365 household benchmark and extended through DexJoCo dexterity, DISCOVERSE simulation, and PAC-MAN whole-body safety control.

Watch demo RoboCasa benchmark

RoboCasa365 Household DexJoCo Dexterity DISCOVERSE Simulation PAC-MAN Safety

16 RoboCasa365 tasks household manipulation benchmark

11 DexJoCo tasks dexterous manipulation benchmark

4-view long-horizon video RoboCasa365 multi-view replay

2 AMD devices 395 + Radeon Pro W7900

Household RoboCasa365

Dexterous DexJoCo

Simulation DISCOVERSE · 3DGS

Safety PAC-MAN

00 / FLAGSHIP

RoboCasa365 official policies on AMD

Official Pi0.5 and GR00T checkpoints complete the same 16-task closed-loop evaluation on the AMD Ryzen AI MAX+ 395. Each model runs 800 formal episodes with per-task results and synchronized four-view video.

PRIMARY SHOWCASE · 4:59 · ENGLISH

AMD Physical AI in motion

From DexJoCo dexterous manipulation and RoboCasa365 household tasks to DISCOVERSE multi-view replay and PAC-MAN safety control, the film shows one AMD workflow across multiple robot capabilities.

Download English MP4 English captions

DexJoCo · Pi0.5 · JAX 0.10DexJoCo dexterous-hand success

DISCOVERSE · AMD · 3-viewHD cabinet-door replay

01 / PROJECTS

Seven complementary evidence tracks

From a teaching baseline to multi-simulator, benchmark, creative-operation, and hardware evidence.

verified result

RoboCasa365 · 16 tasks

Official Pi0.5 and GR00T checkpoints run the same 16-task, 50-episode closed-loop protocol on the AMD 395 with per-task statistics and four-view video.

tasks: 16
episodes: 50 / task
video: 4-view

Open 16-task comparison Long-horizon video

migration complete

DISCOVERSE · full-stack migration

Official scenes, expert trajectories, policy entry points, closed-loop evaluation, HD multi-view replay, and MP4 export form a complete migration path from simulation to visual evidence.

Runtime: 18 / 18
AIRBOT: 12 / 12
Expert data: 500 eps

Migration details 3DGS evidence Upstream

three dodge modes

PAC-MAN · perceptive CBF-RL

The upstream dodge_link_cbf ONNX policy fuses four depth-history frames with G1 proprioception. The robot holds its stance, then autonomously selects a duck, left sidestep, or right sidestep when the ball enters view.

observation: 960-D
modes: 3 / 3
video: 1080p

Open AMD case study Evidence JSON

multi-task evaluated

DexJoCo · Pi0.5 native JAX

Official Orbax weights run natively with AMD ROCm JAX 0.10 across bimanual coordination, tool use, precision grasping, and long-horizon contact tasks, creating a multi-task Pi0.5 dexterous-manipulation showcase.

tasks: 11
runtime: JAX 0.10
media: 10+ videos

Task videos Official weights

device setup

AMD dual-device lab

The local Ryzen AI MAX+ 395 runs sustained evaluation while the cloud Radeon Pro W7900 handles high-memory training and migration validation.

AMD 395: RoboCasa
W7900: VLA + ACT
stack: PyTorch + JAX

Public result JSON, training checkpoints, HD video, and reproduction guides are indexed by workstream.

02 / BENCHMARK

RoboCasa365 matched task benchmark

GR00T 28.75% Pi0.5 17.75%

GR00T N1.5 · official 120k 230 / 800

Pi0.5 · official 75k 142 / 800

Task	GR00T	Pi0.5	Delta
Loading verified results…

same 16 tasks × 50 episodes, video enabled Download result JSON SHA256 76f058…dac94

02B / LONG-HORIZON SHOWCASE

GR00T long-horizon success showcase

GR00T N1.5 multi-task weights on the AMD 395 complete restocking and meal-packing sequences in household scenes, with four synchronized camera views for every trace.

2successful kitchen tasks

20evaluated episodes

4-view1920×1080 · 20 fps

View results and SHA

RestockBowls · 10 episodes

Restock task · successful trace

success

PackIdenticalLunches · 10 episodes

Meal-packing task · successful trace

success

Four synchronized views reveal robot motion, wrist detail, and environment state throughout each long-horizon household task.

02C / DISCOVERSE SHOWCASE

DISCOVERSE HD three-view replay

The same AMD-migrated MMK2 expert state machine; each showcase includes native 1920×1080 cam_0, cam_1, and cam_2 recordings composed into one three-view canvas.

4official tasks

12native camera recordings

1920×1080each native view · 20/25 fps

View results and SHA

DISCOVERSE · MMK2 · 3-view · 1920×1080

Box pick · strict success

success

DISCOVERSE · MMK2 · 3-view · 1920×1080

Drawer open · strict success

success

DISCOVERSE · MMK2 · 3-view · 1920×1080

Cabinet door open · strict success

success

DISCOVERSE · MMK2 · 3-view · 1920×1080

Cup to plate · strict success

success

Three synchronized views capture the workspace, end-effector detail, and environment state on one replay timeline.

02D / 3DGS RENDERER

Real Gaussian Splatting rendering on ROCm

The ROCm Gaussian renderer combines point assets, camera motion, lighting, and robot replay into a visual layer for the migrated scenes.

2ROCm renderer targets

2.36Mloaded Gaussian points

640×360non-empty validation frames

View evidence and SHA

RM2 Gaussian Splatting renderer output — GSRendererMuJoCo · ROCmRM2 · 1,334,537 points
`640×360`

SkyRover Gaussian Splatting renderer output — GSRendererMuJoCo · ROCmSkyRover · 1,026,855 points
`640×360`

DYNAMIC EVIDENCE

Dynamic replay: Franka and UR5e

20 fps · 4 s · native video

3DGS example · FrankaROCm renderer replay

3DGS example · UR5eROCm renderer replay

Two renderer targets, two non-empty validation frames, and two dynamic films demonstrate the ROCm Gaussian Splatting path; the DISCOVERSE expert films connect rendering to robot task execution.

03 / SUCCESS EVIDENCE

Success cases

DexJoCo · Pi0.5 native JAX 0.10 · task replay

water_plant · seed 0 success

success

RoboWits · ACT 100k · W7900 · ego view

06_dominos · episode 1 success · 848×480

success

PAC-MAN · Unitree G1 · upstream ONNX · free-base

In-place whole-body dodging

3 dodge modes

Every Embodied · SmolVLA

Red-cup strict success

success

RoboCasa365 · GR00T · 4-view showcase

CloseFridge success · 1920×1080

success

RoboCasa365 · Pi0.5

DrawerToCounter success

success

DISCOVERSE · Expert

Block bridge · strict replay

success

03 / DEXJOCO MULTI-TASK

Pi0.5 success case archive

11 tasks

Official multi-task weights run on AMD ROCm JAX 0.10 across 11 dexterous manipulation tasks. The page presents task videos covering bimanual coordination, tool use, precision grasping, and long-horizon contact, with the complete task map and result JSON archived alongside them.

Official multi-task suite11 tasks

Bimanual coordination, tool use, precision grasping, and long-contact tasks form the complete task family.

AMD runtimeJAX 0.10

Official Orbax weights run natively on AMD ROCm, with video and JSON produced by one evaluation entry point.

Task media archive10 success traces

The gallery presents bimanual motion, precision contact, tool use, and deterministic success-seed replays.

Task	Capability	Task media
bimanual_assembly	two-hand assembly	View task
bimanual_hanoi	long-horizon placement	View task
bimanual_microwave_cook	tool + contact coordination	View task
bimanual_photograph	precision pose control	View task
click_mouse	fine manipulation	View official demo
fold_glasses	deformable object handling	View official demo
hammer_nail	tool use	View official demo
pick_bucket	grasp and lift	View official demo
pinch_tongs	precision pinch	View official demo
water_plant	pouring and release	View task

TASK VIDEO ARCHIVE

Bimanual coordination, tool use, and precision grasping

official demos + task traces

official · seed 0click_mouse

official · seed 0fold_glasses

official · seed 0hammer_nail

official · seed 0pick_bucket

official · seed 0pinch_tongs

success seed 3bimanual_assembly

success seed 4bimanual_hanoi

success seed 1bimanual_microwave_cook

success seed 3bimanual_photograph

success seed 2water_plant

Task results and runtime archived View result JSON View migration details

04 / PIPELINE

One auditable loop

01SimulationRoboCasa · MuJoCo · DISCOVERSE
02Datafixed seed · LeRobot · audit
03AMD trainingSmolVLA · Pi0.5 · GR00T · ACT
04Closed-loop evaldenominator · success/failure videos
05EvidenceJSON · SHA256 · checkpoint

05 / TRAINING

Training

The public tutorial provides runnable notebooks, protected weights, configs, strict evaluation summaries, and SHA256 records.

Six end-to-end notebooksSmolVLA / Pi0 / ACT, ordinary and protected training SmolVLA protected checkpoint57/60 formal result, configuration, video, and SHA Pi0 advanced checkpoint12/14 strict · unseen 9/10 ACT training checkpointTraining logs, action policy, and physical replay RoboWits checkpoint archiveW7900 100k model and intermediate checkpoints Public integrity manifestSHA-256 for representative media and result JSON AMD migration detailsgates, hard cases, boundaries, and the DexJoCo JAX route

99 / FAILURE CASES

Failure cases

Boundary-case replays expose policy behavior across navigation, approach, contact, grasp, and placement with synchronized visual evidence.

RoboCasa365 · GR00T · 4-view showcase

CloseFridge failure

failure

RoboCasa365 · Pi0.5

DrawerToCounter failure

failure

DISCOVERSE · Pi0.5 diagnostic

Closed-loop policy failure

failure

Every Embodied · SmolVLA

no-lift failure diagnosis

failure