Flattened observations disagree with explicitly declared task specifications
还没有人认领这个 Issue。
评估
- 难度
- 3/5
- 预计耗时
- 1-2 天
- 新手友好度
- 74/100
- Issue 类型
- 缺陷
- 描述清晰度
- 基本清楚
- 活跃度
- 活跃
- 技术栈
- numpy, python
调研方向
Start at control.Environment.observation_spec and compare the explicit-spec path with the existing fallback flattening path used when no specification is declared. Check that flat_observation preserves shapes and NumPy dtype promotion without materializing observations or calling get_observation, while leaving the caller's spec and mapping type unchanged. Run the regression against reset and step behavior to confirm the declared and inferred paths agree.
由索引模型根据 Issue 内容生成。
描述
Reproduction
On current main, an environment using flat_observation=True flattens the values returned by reset and step, but returns an unflattened specification when the task implements observation_spec.
For example, a task declaring a float32 position array of shape (2,) and an int32 scalar mode produces one float64 observations array of shape (3,). Environment.observation_spec() instead returns the original two specifications.
The fallback path for tasks without an explicit specification already flattens correctly. Agents that allocate or validate observations from the environment specification therefore behave differently depending on whether the task supplies its own spec.
Expected behavior
Flatten the declared shapes and apply NumPy's concatenation dtype promotion when flattening is enabled, without materializing observation arrays or calling get_observation. Preserve the caller's spec, mapping type, normal unflattened behavior, and the existing inference path.
Reproduced with actual control.Environment calls and a small native MuJoCo task. The new regression fails on current main at a04e3e4cf56c12117d2294bb090f9acec21e5c67; no renderer, trained policies or external service is required.
- 主要语言
- Python
- 星标
- 4.7k
- 派生
- 768
- PR 合并指标
- 30 天内没有已合并 PR
环境准备
- 没有 Dockerfile 或 Docker Compose 文件
- 没有 Pull Request 模板
- 阅读贡献指南
从这里开始
- 先读完整个 Issue，再读项目的贡献指南。
- 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
- Fork 仓库，在一个分支上完成修改。
- 提交 Pull Request，并在描述里引用这个 Issue 编号。
google-deepmind/dm_control 的其他 Issue
-
难度 2/5 1-3 小时 新手友好度 85/100
google-deepmind/dm_control#559 ·
-
难度 3/5 1-2 天 新手友好度 75/100
google-deepmind/dm_control#560 ·
-
难度 3/5 1-2 天 新手友好度 76/100
google-deepmind/dm_control#552 ·
-
难度 3/5 1-2 天 新手友好度 45/100
google-deepmind/dm_control#540 · 1 条评论 ·
-
难度 3/5 1-2 天 新手友好度 55/100
google-deepmind/dm_control#537 · 1 条评论 ·
查看 google-deepmind/dm_control 的全部 Issue
相似的 Issue
-
namespace operations
难度 1/5 1 小时以内 新手友好度 82/100
EclipseFdn/open-vsx.org#13573 ·
维护者通常 1 天内回复
-
难度 2/5 1-3 小时 新手友好度 72/100
collective/icalendar#1854 ·
维护者通常 1 天内回复
-
难度 2/5 1-3 小时 新手友好度 72/100
rancher/rancher-ai-agent#412 ·
维护者通常 6 天内回复
-
难度 2/5 1-3 小时 新手友好度 84/100
TUDelftGeodesy/DePSI#134 ·
-
难度 2/5 1-3 小时 新手友好度 88/100
HenriquesLab/rxiv-maker#335 ·