Keystroke nATEt

A normalized, time-aligned translational shape error with a translation activity guard:

$$\mathrm{ATE}_t = \frac{\sum_i w_i \cdot (e_{t,\mathrm{start},i} + e_{t,\mathrm{end},i}) / 2}{\sum_i w_i} \tag{1}$$ $$\mathrm{shape}_t = \mathrm{clamp}\!\left( \frac{\mathrm{ATE}_t}{\max(\ell_{\mathrm{obs}},\, \ell_{\min})}, \,0,\,1 \right) \tag{2}$$ $$\mathrm{activity}_t = \frac{n_{\mathrm{mismatch}}}{n_{\mathrm{active}}} \tag{3}$$ $$\mathrm{nATEt} = \max(\mathrm{shape}_t,\, \mathrm{activity}_t) \tag{4}$$

Here $e_{t,\mathrm{start},i}$ and $e_{t,\mathrm{end},i}$ are positional errors against the target at the two ends of observed interval $i$, within its constant-action segment; $w_i = 1 / \mathrm{FPS}$. $\mathrm{ATE}_t$ is the trapezoidal time integral divided by total observed duration, a time-weighted mean error rather than an unnormalized integral. No arc-length resampling is used. The default minimum path normalizer is $\ell_{\min} = 0.5$ pose units. $\ell_{\mathrm{obs}}$ is the observed path length. $n_{\mathrm{mismatch}}$ counts intervals whose commanded and observed translation-activity states disagree; $n_{\mathrm{active}}$ counts intervals with commanded or observed translation.

Observed translation activity means speed exceeds move_speed_threshold=0.06 pose units/second by default. Commanded but stalled motion and uncommanded motion contribute to activity error; observation gaps are missing data, not freezes.

Lower raw error is better. If translation is neither commanded nor observed, the metric is not_applicable, not a perfect score. Target amplitude is prediction-derived (amplitude_calibrated=false); scale and speed accuracy are not independently validated. Backend: pose_action; official metric.

60-Second Same-Input Examples

Audit manifest: docs/assets/metric_examples/longseq_memory_60s_examples_20260909.json.

Input image

60s memory_loop · same benchmark case · gap 0.3023

memory-dolly_loop-memory_pool_277

Suite: memory_loopScene: photorealistic / outdoor / terrestrial_landscape

Benchmark prompt target

Move the camera forward through the scene, then back to where it started.

prompt source: memory_loop_prompt

Compare recovered translation over the 60s path: requested movement, stalls, and extra movement all affect this score. Visual quality is measured separately.

Model result videos will be available with the ModelScope release. / 模型推理视频将在 ModelScope 方案上线后提供。

Higher metric scoreScore 0.8450 · official

Model result videos will be available with the ModelScope release. / 模型推理视频将在 ModelScope 方案上线后提供。

Lower metric scoreScore 0.5427 · official

Input image

60s memory_loop · same benchmark case · gap 0.4043

memory-palindrome_loop-memory_pool_406

Suite: memory_loopScene: photorealistic / outdoor / terrestrial_landscape

Benchmark prompt target

Move forward, turn and continue, then retrace the whole path back.

prompt source: memory_loop_prompt

Compare recovered translation over the 60s path: requested movement, stalls, and extra movement all affect this score. Visual quality is measured separately.

Model result videos will be available with the ModelScope release. / 模型推理视频将在 ModelScope 方案上线后提供。

Higher metric scoreScore 0.7876 · official

Model result videos will be available with the ModelScope release. / 模型推理视频将在 ModelScope 方案上线后提供。

Lower metric scoreScore 0.3834 · official

Input image

60s memory_loop · same benchmark case · gap 0.2480

memory-double_yaw_loop-memory_pool_177

Suite: memory_loopScene: photorealistic / indoor / dining_spaces

Benchmark prompt target

Turn away and back to the start twice in a row.

prompt source: memory_loop_prompt

Compare recovered translation over the 60s path: requested movement, stalls, and extra movement all affect this score. Visual quality is measured separately.

Model result videos will be available with the ModelScope release. / 模型推理视频将在 ModelScope 方案上线后提供。

Higher metric scoreScore 0.8547 · official

Model result videos will be available with the ModelScope release. / 模型推理视频将在 ModelScope 方案上线后提供。

Lower metric scoreScore 0.6067 · official

Input image

60s memory_loop · same benchmark case · gap 0.3874

memory-dolly_loop-memory_pool_065

Suite: memory_loopScene: photorealistic / outdoor / terrestrial_landscape

Benchmark prompt target

Move the camera forward through the scene, then back to where it started.

prompt source: memory_loop_prompt

Compare recovered translation over the 60s path: requested movement, stalls, and extra movement all affect this score. Visual quality is measured separately.

Model result videos will be available with the ModelScope release. / 模型推理视频将在 ModelScope 方案上线后提供。

Higher metric scoreScore 0.8406 · official

Model result videos will be available with the ModelScope release. / 模型推理视频将在 ModelScope 方案上线后提供。

Lower metric scoreScore 0.4532 · official

Input image

60s memory_loop · same benchmark case · gap 0.3420

memory-yaw_loop_360-memory_pool_416

Suite: memory_loopScene: photorealistic / outdoor / terrestrial_landscape

Benchmark prompt target

Rotate the camera a full turn in place and come back to the starting view.

prompt source: memory_loop_prompt

Compare recovered translation over the 60s path: requested movement, stalls, and extra movement all affect this score. Visual quality is measured separately.

Model result videos will be available with the ModelScope release. / 模型推理视频将在 ModelScope 方案上线后提供。

Higher metric scoreScore 0.8744 · official

Model result videos will be available with the ModelScope release. / 模型推理视频将在 ModelScope 方案上线后提供。

Lower metric scoreScore 0.5324 · official