refactor!: state.info ownership — episode state on env buffers, typed reward_terms, reset -> None

docs/source/en/user_guide/envs/humanoid_velocity_tracking/env_design.md

@@ -73,7 +73,7 @@ sole every step, so clearance remains relative to the local surface.
 Each raw term is multiplied by its `RewardScales` weight and the control timestep; negative weights turn constraint
 measurements such as `penalty_*` and `pose` into penalties. Curriculum-selected penalties are also multiplied by the
 current `penalty_scale` according to completed episode length. This factor is exposed through
-`info["metrics"]["penalty_scale"]`, and the final weighted terms through `info["Reward"]`.
+`state.metrics["penalty_scale"]`, and the final weighted terms through `state.reward_terms`.
 
 ## Termination conditions
 

docs/source/en/user_guide/envs/quadruped_velocity_tracking/config_tuning.md

@@ -128,7 +128,7 @@ reward_config = RewardConfig(
 
 `base_height_target` and `initial_base_position[2]` should normally be close to the robot's default standing height.
 Choose `target_foot_height` according to leg length and terrain variation: too little encourages dragging, while too much
-can require unreasonable joint motion. When tuning a reward weight, inspect the matching value in `info["Reward"]`
+can require unreasonable joint motion. When tuning a reward weight, inspect the matching value in `state.reward_terms`
 rather than comparing configuration numbers alone.
 
 ## 7. `sensor`: sensor-name mapping

docs/source/en/user_guide/envs/quadruped_velocity_tracking/env_design.md

@@ -79,7 +79,7 @@ a body reference frame. Contact matching, swing clearance, and early swing-conta
 | `swing_contact`      | Measure the fraction of feet still touching the ground during swing                                  | Penalize dragging and early touchdown       |
 
 Each raw term is first multiplied by its `RewardScales` weight; negative weights turn non-negative measurements into
-penalties. `info["Reward"]` stores these weighted values before timestep scaling. Their sum is multiplied by `ctrl_dt` to
+penalties. `state.reward_terms` stores these weighted values before timestep scaling. Their sum is multiplied by `ctrl_dt` to
 produce the reward returned to the training algorithm. The current environment has no reward curriculum.
 
 On rough terrain, `base_height` uses terrain height beneath the robot as its zero point. `swing_feet_z` consumes foot

docs/source/en/user_guide/envs/whole_body_tracking/adding_wbt_task.md

@@ -139,7 +139,7 @@ python scripts/train.py task=g1-wbt-dance1-subject1/motrix.fastsac \
 ```
 
 Check for missing motion, joint, or body names; systematic NaNs, joint-limit violations, or immediate bad-tracking after
-reset; and verify that `info["Reward"]` and `info["metrics"]` reach the logs. Then start the default training run:
+reset; and verify that `state.reward_terms` and `state.metrics` reach the logs. Then start the default training run:
 
 ```bash
 python scripts/train.py task=g1-wbt-dance1-subject1/motrix.fastsac algo.asynchronous=true

docs/source/en/user_guide/envs/whole_body_tracking/env_design.md

@@ -81,7 +81,7 @@ This makes the relative body-configuration reward independent of the current hor
 | `undesired_contacts`                         | Count robot links whose net contact force exceeds the threshold and are absent from `allowed_contact_links` | Suppress body contacts not required by the motion    |
 
 Each raw term is multiplied by its `WbtRewardScales` weight and then by `ctrl_dt`. Negative weights turn `action_rate_l2`,
-`limits_dof_pos`, and `undesired_contacts` into penalties. Final weighted terms are written to `info["Reward"]`.
+`limits_dof_pos`, and `undesired_contacts` into penalties. Final weighted terms are written to `state.reward_terms`.
 
 ## Termination conditions
 
@@ -96,7 +96,7 @@ Each raw term is multiplied by its `WbtRewardScales` weight and then by `ctrl_dt
 | Motion final frame          | Neither      | `motion_steps` reaches the clip end                                                                                    | Resample and reset motion state during training; restart at frame 0 during play |
 
 `undesired_contacts` contributes only a reward penalty and does not terminate the episode. Termination rates and error means
-are written to `info["metrics"]`.
+are written to `state.metrics`.
 
 ## Reset logic
 

docs/source/en/user_guide/envs/whole_body_tracking/index.md

@@ -160,8 +160,8 @@ python scripts/train.py task=g1-wbt-dance/motrix.fastsac \
 ```
 
 During training, environments start at different points in the motion, and failure records increase the sampling probability
-of difficult regions. Weighted reward terms are written to `info["Reward"]`; bad-tracking rates, motion progress, and
-adaptive-sampling statistics are written to `info["metrics"]`. See [Task Environment Design](env_design.md) for definitions.
+of difficult regions. Weighted reward terms are written to `state.reward_terms`; bad-tracking rates, motion progress, and
+adaptive-sampling statistics are written to `state.metrics`. See [Task Environment Design](env_design.md) for definitions.
 
 ### Play the policy
 

docs/source/en/user_guide/tutorial/building_envs/direct_env.md

@@ -146,7 +146,7 @@ Hooks a subclass implements:
 
 | Hook                                        | Responsibility                                                                        |
 | ------------------------------------------- | -------------------------------------------------------------------------------------- |
-| `reset(env_ids)`                            | Write reset state (randomized initial poses, ...) for the selected rows and return an info dict; observations are produced afterwards by `compute_observation` |
+| `reset(env_ids)`                            | Write reset state (randomized initial poses, ...) for the selected rows; observations are produced afterwards by `compute_observation` |
 | `apply_action(actions, state)`              | Write the action into the simulator (usually ctrl targets)                             |
 | `compute_transition(state)`                 | Execute the read program and derive `state.reward`, `state.terminated`, ...; this is the only full data refresh of a step and must **not** write `state.obs` |
 | `compute_observation(state)`                | Assemble `state.obs` purely from already refreshed simulator data, without further reads |
@@ -156,7 +156,7 @@ Semantics:
 
 - `terminated` marks episode-ending conditions such as task failure; `truncated` marks
   the time limit at `max_episode_steps`. `ArrayEnv` combines both into `done` and
-  triggers auto-reset; `info["time_outs"]` flags rows that timed out without failing.
+  triggers auto-reset.
 - The environment dimension must use vectorized NumPy operations; plain loops are only
   allowed over a fixed number of joints, feet, or terms.
 - Constants (initial poses, space definitions, query names) are precomputed in

docs/source/en/user_guide/tutorial/building_envs/index.md

@@ -34,7 +34,7 @@ Every stage is orchestrated by the `ArrayEnv` base class; environment implementa
 hooks and must not re-implement the lifecycle. Semantics:
 
 - `terminated` marks episode-ending conditions such as task failure; `truncated` marks
-  the time limit, and `info["time_outs"]` flags rows that timed out without failing;
+  the time limit;
 - environments that are done are reset automatically at the end of each step, and
   observations are recomputed after the reset.
 

docs/source/en/user_guide/tutorial/training/training_and_result.md

@@ -95,7 +95,7 @@ TensorBoard logs are written under the run directory and can be viewed per envir
 tensorboard --logdir runs/cartpole
 ```
 
-Besides the standard return and loss curves, if an environment exposes per-term rewards via `info["Reward"]`, they are also logged to TensorBoard during training.
+Besides the standard return and loss curves, if an environment exposes per-term rewards via `state.reward_terms`, they are also logged to TensorBoard during training.
 
 ## Model Evaluation and Testing
 

docs/source/zh_CN/user_guide/envs/humanoid_velocity_tracking/env_design.md

@@ -66,7 +66,7 @@ $$
 
 每个原始项先乘以 `RewardScales` 中的权重，再乘以控制步长；`penalty_*` 和 `pose` 等约束项通过负权重成为惩罚。
 课程指定的惩罚项还会根据已结束回合的平均长度乘以当前 `penalty_scale`。该缩放记录在
-`info["metrics"]["penalty_scale"]`，最终的各加权项记录在 `info["Reward"]`。
+`state.metrics["penalty_scale"]`，最终的各加权项记录在 `state.reward_terms`。
 
 ## 终止条件
 

docs/source/zh_CN/user_guide/envs/quadruped_velocity_tracking/config_tuning.md

@@ -122,7 +122,7 @@ reward_config = RewardConfig(
 
 `base_height_target` 和 `initial_base_position[2]` 通常应接近机器人的默认站立高度。`target_foot_height` 应结合
 腿长和地形起伏设置；过低容易拖脚，过高可能要求超出合理关节范围。调节某个奖励权重时，应查看
-`info["Reward"]` 中对应项的量级，而不只比较配置数值。
+`state.reward_terms` 中对应项的量级，而不只比较配置数值。
 
 ## 7. `sensor`：传感器名称映射
 

docs/source/zh_CN/user_guide/envs/quadruped_velocity_tracking/env_design.md

@@ -73,7 +73,7 @@ Actor 与 Critic 共享的字段使用同一份带噪值；当前实现没有额
 | `swing_feet_z`       | 在摆动且未接触时，对足端高度误差应用指数核并按脚数平均 | 鼓励摆动脚达到目标离地高度 |
 | `swing_contact`      | 计算摆动期仍接触地面的脚所占比例                       | 惩罚拖脚和过早落足         |
 
-每个原始项先乘以 `RewardScales` 中的权重；负权重将非负度量转为惩罚。`info["Reward"]` 保存这些尚未
+每个原始项先乘以 `RewardScales` 中的权重；负权重将非负度量转为惩罚。`state.reward_terms` 保存这些尚未
 乘控制步长的加权项，所有项求和后再乘 `ctrl_dt`，得到返回给训练算法的单步奖励。当前环境没有奖励课程。
 
 粗糙地形上，`base_height` 使用机器人当前位置的地形高度作为零点。`swing_feet_z` 的足端位置来自机体参考系，

docs/source/zh_CN/user_guide/envs/whole_body_tracking/adding_wbt_task.md

@@ -136,7 +136,7 @@ python scripts/train.py task=g1-wbt-dance1-subject1/motrix.fastsac \
 ```
 
 检查 motion/joint/body 名称没有缺失，reset 后没有系统性 NaN、joint limit 违规或立即 bad-tracking，并确认
-`info["Reward"]` 与 `info["metrics"]` 能进入日志。随后使用默认规模训练：
+`state.reward_terms` 与 `state.metrics` 能进入日志。随后使用默认规模训练：
 
 ```bash
 python scripts/train.py task=g1-wbt-dance1-subject1/motrix.fastsac algo.asynchronous=true

docs/source/zh_CN/user_guide/envs/whole_body_tracking/env_design.md

@@ -77,7 +77,7 @@ Actor 的参考身体姿态、基座角速度、关节位置和关节速度会
 | `undesired_contacts`                         | 统计净接触力超过阈值且不在 `allowed_contact_links` 中的机器人 links | 抑制动作不需要的身体接触         |
 
 每个原始项先乘以 `WbtRewardScales` 中的权重，再乘以 `ctrl_dt`；`action_rate_l2`、`limits_dof_pos` 和
-`undesired_contacts` 通过负权重成为惩罚。最终加权项写入 `info["Reward"]`。
+`undesired_contacts` 通过负权重成为惩罚。最终加权项写入 `state.reward_terms`。
 
 ## 终止条件
 
@@ -91,7 +91,7 @@ Actor 的参考身体姿态、基座角速度、关节位置和关节速度会
 | 时间上限         | `truncated`  | 训练回合达到 `max_episode_seconds`，内置配置为 10 s                                | 正常达到训练时限，不表示 bad tracking                      |
 | Motion 末帧      | 两者都不是   | `motion_steps` 到达 clip 末尾                                                      | 训练时重采样起始帧并重置 motion 状态；play 时从第 0 帧重播 |
 
-`undesired_contacts` 只产生奖励惩罚，不直接终止回合。各类终止比例与误差均值记录在 `info["metrics"]`。
+`undesired_contacts` 只产生奖励惩罚，不直接终止回合。各类终止比例与误差均值记录在 `state.metrics`。
 
 ## 重置逻辑
 

docs/source/zh_CN/user_guide/envs/whole_body_tracking/index.md

@@ -154,7 +154,7 @@ python scripts/train.py task=g1-wbt-dance/motrix.fastsac \
 ```
 
 训练期间，环境从 motion 的不同时间点开始，并利用失败记录提高困难片段的采样概率。奖励分项写入
-`info["Reward"]`，bad-tracking 比例、motion 进度和自适应采样统计写入 `info["metrics"]`；各项定义见
+`state.reward_terms`，bad-tracking 比例、motion 进度和自适应采样统计写入 `state.metrics`；各项定义见
 [任务环境设计](env_design.md)。
 
 ### 回放策略

docs/source/zh_CN/user_guide/tutorial/building_envs/direct_env.md

@@ -139,7 +139,7 @@ class CartPoleEnv(DirectEnv):
 
 | 钩子                                 | 职责                                                                                    |
 | ------------------------------------ | ---------------------------------------------------------------------------------------- |
-| `reset(env_ids)`                     | 为选中的环境写入重置状态（随机化初始姿态等），返回 info 字典；观测交给后续的 `compute_observation` |
+| `reset(env_ids)`                     | 为选中的环境写入重置状态（随机化初始姿态等）；观测交给后续的 `compute_observation` |
 | `apply_action(actions, state)`       | 将动作写入仿真（通常是 ctrl 目标）                                                       |
 | `compute_transition(state)`          | 执行读取程序并派生 `state.reward`、`state.terminated` 等；这是每步唯一的全量数据刷新点，**不得**写 `state.obs` |
 | `compute_observation(state)`         | 纯粹用已刷新的仿真数据拼装 `state.obs`，不再执行读取                                     |
@@ -148,8 +148,7 @@ class CartPoleEnv(DirectEnv):
 语义约定：
 
 - `terminated` 表示任务失败等回合终止条件；`truncated` 表示达到 `max_episode_steps`
-  的时间截断。两者由 `ArrayEnv` 合成 `done` 并触发 auto-reset；
-  `info["time_outs"]` 标记"截断但未失败"的行。
+  的时间截断。两者由 `ArrayEnv` 合成 `done` 并触发 auto-reset。
 - 环境维度必须使用 NumPy 向量化操作；只有遍历固定数量的关节、脚或 term 时才允许普通循环。
 - 常量（初始姿态、空间定义、query 名称）在 `__init__` 或配置构造阶段预计算，
   不在 step 循环中重复创建。

docs/source/zh_CN/user_guide/tutorial/building_envs/index.md

@@ -28,8 +28,7 @@
 
 所有阶段都由 `ArrayEnv` 基类编排，环境实现只填充钩子，不要重复实现生命周期。语义约定：
 
-- `terminated` 是任务失败等回合终止条件；`truncated` 是达到回合时长上限的时间截断，
-  `info["time_outs"]` 标记"截断但未失败"的行；
+- `terminated` 是任务失败等回合终止条件；`truncated` 是达到回合时长上限的时间截断；
 - done 的环境在每步末尾被自动重置，观测在重置之后重新计算。
 
 各阶段在两种工作流中分别由谁实现：DirectEnv 在

docs/source/zh_CN/user_guide/tutorial/training/training_and_result.md

@@ -95,7 +95,7 @@ TensorBoard 日志写在 run 目录下，可按环境查看：
 tensorboard --logdir runs/cartpole
 ```
 
-除标准的回报、损失曲线外，若环境通过 `info["Reward"]` 暴露了各 reward 分项，训练时也会将其记录到 TensorBoard。
+除标准的回报、损失曲线外，若环境通过 `state.reward_terms` 暴露了各 reward 分项，训练时也会将其记录到 TensorBoard。
 
 ## 模型评估和测试
 

motrix_deploy_tasks/tests/test_training_parity.py

@@ -27,7 +27,7 @@ def test_go2_training_and_deployment_task_golden_probe(env_name: str) -> None:
     context = PolicyContext(
         step=0,
         elapsed_time_s=0.0,
-        command=PlanarVelocityCommand(env_state.info["commands"]),
+        command=PlanarVelocityCommand(env._commands),
     )
     task = create_task(profile.task, profile.robot)
     command_scale = np.asarray(profile.task.config["command_scale"], dtype=np.float32)
@@ -53,7 +53,7 @@ def test_go2_training_and_deployment_task_golden_probe(env_name: str) -> None:
     next_context = PolicyContext(
         step=1,
         elapsed_time_s=profile.control.period_s,
-        command=PlanarVelocityCommand(stepped.info["commands"]),
+        command=PlanarVelocityCommand(env._commands),
     )
     np.testing.assert_allclose(
         task.build_observation(next_state, next_context),

motrix_env_core/src/motrix_env_core/array/env.py

@@ -40,7 +40,10 @@ class ArrayEnvState:
     terminated: np.ndarray
     truncated: np.ndarray
     episode_steps: np.ndarray
-    info: dict
+    # Per-term reward breakdown for the latest transition: term name to a
+    # ``(num_envs,)`` array. Freshly written every transition; empty when the
+    # environment does not decompose its reward.
+    reward_terms: dict[str, np.ndarray] = dataclasses.field(default_factory=dict)
     # Live diagnostics view: per-environment quantities stay unreduced as
     # ``(num_envs,)`` arrays and are views into manager buffers that kernels
     # overwrite every step; batch-level gauges are scalars. Values always
@@ -150,8 +153,7 @@ def init_state(self) -> ArrayEnvState:
         terminated = np.ones((self._num_envs,), dtype=bool)
         truncated = np.zeros((self._num_envs,), dtype=bool)
         episode_steps = np.zeros((self._num_envs,), dtype=np.uint64)
-        info = {"time_outs": np.zeros((self._num_envs,), dtype=bool)}
-        self._state = self._new_state(obs, reward, terminated, truncated, episode_steps, info)
+        self._state = self._new_state(obs, reward, terminated, truncated, episode_steps)
         self._reset_done_envs()
         with self.perf.scope("observation"):
             self._state = self.compute_observation(self._state)
@@ -166,7 +168,6 @@ def _new_state(
         terminated: np.ndarray,
         truncated: np.ndarray,
         episode_steps: np.ndarray,
-        info: dict,
     ) -> ArrayEnvState:
         """Assemble the environment state; subclasses add simulator-owned fields."""
         return ArrayEnvState(
@@ -175,7 +176,6 @@ def _new_state(
             terminated=terminated,
             truncated=truncated,
             episode_steps=episode_steps,
-            info=info,
         )
 
     @property
@@ -223,31 +223,11 @@ def _reset_done_envs(self) -> None:
             np.putmask(state.episode_steps, done, 0)
             env_ids = np.flatnonzero(done)
         with self.perf.scope("reset_envs"):
-            info1 = self.reset(env_ids)
-        self._merge_reset_info(state, info1, done)
-
-    def _merge_reset_info(self, state: ArrayEnvState, info1: dict, done: np.ndarray) -> None:
-        """Merge one selected-row reset's info entries into the state info."""
-        if not info1:
-            return
-
-        def replace_dict_values(dst, new_values, mask):
-            for key, value in new_values.items():
-                if key not in dst:
-                    dst[key] = value
-                else:
-                    if isinstance(value, np.ndarray):
-                        dst[key][mask] = value
-                    elif isinstance(value, dict):
-                        assert isinstance(dst[key], dict)
-                        replace_dict_values(dst[key], value, mask)
-
-        with self.perf.scope("merge_info"):
-            replace_dict_values(state.info, info1, done)
+            self.reset(env_ids)
 
     @abc.abstractmethod
-    def reset(self, env_ids: np.ndarray) -> dict:
-        """Write reset rows for the selected environments and return reset info.
+    def reset(self, env_ids: np.ndarray) -> None:
+        """Write reset rows for the selected environments.
 
         Subclasses address their simulator's rows themselves. Observation
         generation is deferred until :meth:`compute_observation`, after reset
@@ -265,7 +245,6 @@ def _update_truncate(self):
         if not max_episode_steps:
             return
         self._state.truncated = self._state.episode_steps >= max_episode_steps
-        self._state.info["time_outs"] = self._state.truncated & ~self._state.terminated
 
     @abc.abstractmethod
     def create_renderer(self, config: RenderConfig) -> SimRenderer: