1 00:00:00,000 --> 00:00:04,288 Datawhale EAI builds a reproducible AMD Physical AI stack 2 00:00:04,438 --> 00:00:09,780 across dexterous hands, household manipulation, simulation, and rendering. 3 00:00:10,523 --> 00:00:13,025 The film begins with a real task outcome, 4 00:00:13,451 --> 00:00:17,493 then follows the evidence back to the code and the machine that produced it. 5 00:00:17,990 --> 00:00:21,915 One evidence contract follows data, policy, physics, and proof 6 00:00:21,915 --> 00:00:23,799 across the complete project. 7 00:00:24,475 --> 00:00:27,583 Every result keeps its task definition, model, seed, environment, 8 00:00:27,716 --> 00:00:30,044 video, metrics, and checksum together, 9 00:00:30,044 --> 00:00:35,441 so the viewer can move from a beautiful frame to a reproducible run. 10 00:00:36,960 --> 00:00:42,194 RoboCasa three sixty-five gives us a household manipulation benchmark 11 00:00:42,460 --> 00:00:45,521 with fixed tasks, seeds, videos, and a mobile observation contract. 12 00:00:46,256 --> 00:00:49,215 The kitchen scenes make the target concrete: 13 00:00:49,390 --> 00:00:52,622 reach, grasp, transport, open, place, and recover. 14 00:00:53,248 --> 00:00:57,056 The official task boundary stays visible throughout every household trace. 15 00:00:59,090 --> 00:01:03,219 Long-horizon success traces make the full action sequence inspectable. 16 00:01:03,877 --> 00:01:07,083 Packing and restocking show why the trajectory matters: 17 00:01:07,593 --> 00:01:10,625 the robot must keep contact, preserve the object, 18 00:01:10,927 --> 00:01:14,304 complete several sub-actions, and finish in the correct state. 19 00:01:14,963 --> 00:01:19,655 Shorter attempts move quickly to the next evidence clip, keeping the story focused. 20 00:01:21,730 --> 00:01:25,952 DexJoCo brings contact-rich dexterous tasks into the same protocol, 21 00:01:26,405 --> 00:01:28,849 with native JAX running on AMD ROCm. 22 00:01:28,849 --> 00:01:32,195 Bimanual assembly, Hanoi, and microwave manipulation 23 00:01:32,195 --> 00:01:35,743 put the hand in the difficult part of the problem: 24 00:01:36,286 --> 00:01:40,240 precise contact, coordinated fingers, and long action sequences. 25 00:01:40,240 --> 00:01:46,647 These are the success traces worth giving time to, because the outcome is easy to inspect. 26 00:01:47,150 --> 00:01:50,230 Each result carries its own protocol and denominator. 27 00:01:50,880 --> 00:01:52,642 Official seeds set the benchmark, 28 00:01:52,964 --> 00:01:56,081 recovery searches answer a separate engineering question, 29 00:01:56,445 --> 00:02:00,256 and single-task diagnostics explain a specific behavior. 30 00:02:00,898 --> 00:02:05,361 The film keeps every claim attached to its source while moving at a steady pace. 31 00:02:05,987 --> 00:02:08,867 Every line remains ready for review and reuse. 32 00:02:11,740 --> 00:02:13,332 The evidence contract is simple: 33 00:02:13,753 --> 00:02:18,294 data becomes policy, policy meets physics, and every result becomes proof. 34 00:02:18,955 --> 00:02:21,684 Demonstrations are converted into a standard dataset, 35 00:02:22,113 --> 00:02:25,050 the policy is trained or fine-tuned on AMD, 36 00:02:25,488 --> 00:02:27,361 the simulator executes the action loop, 37 00:02:27,726 --> 00:02:29,850 and the evaluator records stage outcomes. 38 00:02:30,566 --> 00:02:36,145 The final package joins code, a command, a video, a JSON record, and a hash. 39 00:02:37,290 --> 00:02:40,545 DISCOVERSE preserves expert paths, policy experiments, 40 00:02:40,863 --> 00:02:43,913 multi-view rendering, and MP4 output after migration. 41 00:02:43,913 --> 00:02:47,740 The same scene can be inspected from the task camera, 42 00:02:48,026 --> 00:02:50,497 an external view, and a diagnostic angle. 43 00:02:51,201 --> 00:02:52,709 That makes migration visible: 44 00:02:53,128 --> 00:02:54,772 the environment still resets, 45 00:02:55,100 --> 00:02:56,124 the expert path still runs, 46 00:02:56,124 --> 00:02:58,676 the policy loop still produces evidence, 47 00:02:58,676 --> 00:03:01,781 and the video remains useful to a learner. 48 00:03:02,280 --> 00:03:05,299 Three-D Gaussian Splatting and dynamic replay 49 00:03:05,299 --> 00:03:09,109 turn the renderer itself into visible migration evidence. 50 00:03:09,109 --> 00:03:11,274 Rendering is a first-class AMD path: 51 00:03:11,889 --> 00:03:14,755 point assets, camera motion, lighting, camera path, 52 00:03:15,038 --> 00:03:19,390 and replay output travel through the same reproducibility checks 53 00:03:19,390 --> 00:03:21,331 as the policy, with detail. 54 00:03:23,690 --> 00:03:27,270 ROCm versions, GPU identity, memory, throughput, training time, 55 00:03:27,449 --> 00:03:30,858 JSON, MP4, and SHA travel with every result. 56 00:03:30,858 --> 00:03:33,813 We record the execution path explicitly: 57 00:03:33,813 --> 00:03:37,656 PyTorch or JAX, the HIP runtime, the selected GPU, 58 00:03:37,656 --> 00:03:40,785 the memory footprint, the training command, 59 00:03:41,014 --> 00:03:41,910 the evaluation command, 60 00:03:42,032 --> 00:03:44,306 and the exact artifact hash. 61 00:03:44,306 --> 00:03:49,183 That is how an AMD port becomes evidence rather than a claim. 62 00:03:49,680 --> 00:03:51,067 Success leads the story, 63 00:03:51,360 --> 00:03:53,700 but the source evidence stays attached behind it: 64 00:03:54,203 --> 00:03:56,498 code, videos, metrics, and checksums. 65 00:03:57,148 --> 00:03:59,811 The strongest clips earn a complete moment on screen. 66 00:04:00,389 --> 00:04:02,855 Shorter clips answer one question and move on. 67 00:04:03,474 --> 00:04:05,571 The website then opens the deeper layer: 68 00:04:05,920 --> 00:04:09,981 notebooks, weights, raw evaluation records, environment notes, 69 00:04:10,265 --> 00:04:12,869 and the commands used to reproduce each result. 70 00:04:13,370 --> 00:04:18,475 On AMD, PAC-MAN predictive control turns perception into a whole-body dodge. 71 00:04:19,086 --> 00:04:21,187 Unitree G1 tracks the projectile, predicts its path, 72 00:04:21,503 --> 00:04:25,131 and coordinates its whole body to preserve clearance 73 00:04:25,363 --> 00:04:28,530 as the threat crosses the scene in real time on AMD. 74 00:04:29,168 --> 00:04:34,534 The closing traces then map navigation, contact, release, and recovery into targets. 75 00:04:35,156 --> 00:04:38,757 Each stage stays linked to the video and the next control decision. 76 00:04:39,130 --> 00:04:41,827 Build it, run it, and show the proof 77 00:04:42,047 --> 00:04:46,002 through code, notebooks, weights, reports, videos, and reproducible commands. 78 00:04:46,748 --> 00:04:51,010 Datawhale-EAI brings the result to AMD ROCm as an open, inspectable project: 79 00:04:51,010 --> 00:04:55,881 the next person should be able to run the command, see the motion, 80 00:04:56,156 --> 00:04:58,565 and understand the success and the next step.