HoloEngine on RoboDojo: 78.28, interim results

HiHolo has shared an interim simulation report for the dual-arm ARX X5. Here is what the local CPU-only evaluation shows, how the system acts, and where the comparison stops.

· HiHolo · 5 min read

Illustration: sorting rubber ducks by size. Two robot arms on a table with rubber ducks lined up by size.
Illustration: sorting rubber ducks by size. Local CPU-only MuJoCo simulation, not an official RoboDojo evaluation.

A snapshot of simulated manipulation

A robot manipulation score is most useful when you can inspect how the system reached it. HiHolo's RoboDojo report, dated 6 November 2026, brings together a comparison chart, task scores, recorded attempts, and the workflow behind the run.

The experiment uses a dual-arm ARX X5 in HiHolo's CPU-only MuJoCo replica. It uses RoboDojo's public layouts, assets, and scoring code, but the physics and rendered camera views come from the replica. HoloLanguage selects actions, with recognition support inside HoloEngine.

The snapshot covers 184 episodes across 52 tested configurations, using seed 0 and 3-10 attempts per configuration. The page lists all 42 tasks and averages standard and randomized variants where present. Cloth is absent, so both clothes-folding variants were omitted and the task receives zero.

Coverage is uneven. Averaged over variants, 25 of the 42 tasks receive full credit and 36 score above zero. Six score zero: fold_clothes (omitted), make_toast, pour_by_language, pour_liquid_into_cup, store_laptop_and_headphones and sweep_blocks. Several partial scores are low, including play_stacking_toy at 6.7, hang_mugs at 14.2, fill_egg_holder at 20.0 and store_tools_in_toolbox at 41.7. Counted per tested configuration, 29 receive full credit, 14 partial credit and 9 zero out of 52.

The comparison, with its limits

The reported overall score is 78.28 in the local MuJoCo replica. The calculation weights five benchmark dimensions equally. The table preserves the public page's descending comparison order and selected entries, with the simulation context made explicit.

Comparison reproduced from the public report. The HoloEngine row is a local MuJoCo replica result; all other rows are selected official Isaac Sim leaderboard entries. These are different simulation environments, not an apples-to-apples ranking. An official result requires evaluation by RoboDojo. Local run: MacBook Pro, Intel Core i9-9880H, CPU only; seed 0; 184 episodes across 52 tested configurations; clothes-folding omitted and scored zero.
SystemScoreEvaluation context
HoloEngine (local MuJoCo, CPU only)78.28Local MuJoCo replica, interim
Q0-Preview74.20Official Isaac Sim entry
GO-3 Nano Preview61.90Official Isaac Sim entry
PyRUA-Lean55.26Official Isaac Sim entry
HexaShell-Preview53.18Official Isaac Sim entry
WALL-OSS-1.0-lite42.59Official Isaac Sim entry
GPT-6-Astra28.97Official Isaac Sim entry
Pi-0511.44Official Isaac Sim entry
DeepSeek-Flash2.99Official Isaac Sim entry
GPT-5.51.13Official Isaac Sim entry

The official entries provide context, not a controlled comparison against HiHolo's local run. A higher number in this table does not establish that HoloEngine leads the official RoboDojo benchmark. An official result requires evaluation by RoboDojo.

Sources: HiHolo's interim report and the linked official RoboDojo simulation leaderboard. The table shows nine of the 66 entries on the official simulation leaderboard (as of 3 November 2026), as selected on HiHolo's page. It is a snapshot, not the complete leaderboard.

From task examples to the next motion

Before evaluation, HiHolo prepares task knowledge from expert demonstrations, public layouts, and 3D object models. The preparation step measures grasp and release examples and renders object views and silhouettes. It saves motion examples, task parameters, and visual references for use during an attempt.

Before evaluation: RoboDojo inputs pass through extraction and rendering into task knowledge. Task knowledge feeds Python perception and HoloEngine. During each attempt: the MuJoCo scene feeds Python perception, HoloEngine selects a skill, and motion skills update the scene. Fresh views and grip checks feed the next decision. BEFORE EVALUATION RoboDojo inputs Expert demonstrations, public layouts, and 3D object models Extract and render Grasp and release examples; object views and silhouettes Task knowledge Motion examples, task parameters, and visual references Feeds Python perception and HoloEngine. DURING EACH ATTEMPT MuJoCo scene RGB views, arm poses, gripper opening, and task instruction Python perception Extract visual features and estimate geometry from camera views HoloEngine Recognition matches objects to references. HoloLanguage tracks progress and selects a skill. Function Tree fits and evaluates movement rules. Motion skills Arm and gripper trajectories change the scene. Fresh camera views and grip checks feed the next decision. Simulation workflow adapted from the HiHolo RoboDojo report.
Architecture adapted from the public RoboDojo report. Preparation feeds task knowledge into the local MuJoCo simulation workflow; the return path represents scene updates, fresh views, and grip checks.

During an attempt, the MuJoCo scene supplies RGB views, arm poses, gripper opening, and the instruction. Python perception extracts visual features and estimates object geometry. Recognition in HoloEngine identifies objects from reference examples, while HoloLanguage tracks progress and chooses the next skill. Function Tree fits and evaluates movement rules for selected skills.

Motion skills then build arm and gripper trajectories. The motion changes the scene, and fresh camera views and grip checks feed the next decision. The diagram adapts this sequence from the report; it does not describe a separate architecture.

CPU measurements and open limits

The local evaluation hardware is a MacBook Pro with an Intel Core i9-9880H, using the CPU only. HoloLanguage decisions average 431 ms across 2,397 logged decisions in 52 configurations. That timing includes Function Tree when it runs inside a decision. Separately, fitting and evaluating the block-stacking rule averages 23 ms across 30 calls.

Separate recognition measurements average 1.56 s for two digits, 5.19 s for five objects, and 21.66 s for a text description matched against the full reference library. Each figure averages three requests. These measurements describe different operations, not one interchangeable end-to-end latency.

The report also lists 88.1 MB peak resident memory for the HoloEngine process across seven sampled tasks on CPU, with a fresh process and one episode for each sampled task. These runs were separate from the 184-episode evaluation. That is a process-level measurement, not the memory footprint of the whole simulation: it excludes MuJoCo, rendering, OpenCV and the Python adapter.

Several limits remain. The replica lacks cloth and fluid simulation and a third arm. The sorting demonstration follows recorded order and grasp timing. Tic-tac-toe adds the opponent's mark after the original step count, and opponent play in mahjong is incomplete. Recognition uses visual references from replica images; transfer to Isaac Sim imagery remains unverified.

Read the full RoboDojo benchmark report for per-task scores, recorded runs, and the complete method notes. These interim simulation results document the tested policy and layouts. They do not establish an official RoboDojo result or real-robot performance.

Back to HiHolo insights