Datasets:

Datawhale
/

atec2026-task-e-reproducibility

import torch
import torch.nn as nn
from collections import deque
import torchvision.transforms.functional as TF
import torchvision.transforms as T
from dataclasses import dataclass
import sys
import os

current_path = os.path.dirname(os.path.abspath(__file__))
#sys.path.insert(0, current_path)

from act.detr.backbone import build_backbone
from act.detr.transformer import build_transformer
from act.detr.detr_vae import build_encoder, DETRVAE

@dataclass
class Args:
    torch_deterministic: bool = True
    """if toggled, `torch.backends.cudnn.deterministic=False`"""
    cuda: bool = True
    """if toggled, cuda will be enabled by default"""
    temporal_agg: bool = True
    """if toggled, temporal ensembling will be performed at inference"""

    # Backbone
    position_embedding: str = 'sine'
    backbone: str = 'resnet18'
    lr_backbone: float = 1e-5
    masks: bool = False
    dilation: bool = False
    include_depth: bool = False
    """always False — depth not collected; kept for backbone API compatibility"""
    include_rgb: bool = True
    """use RGB images as input (requires --save_images during collection)"""

    # Transformer
    enc_layers: int = 2
    dec_layers: int = 4
    dim_feedforward: int = 512
    hidden_dim: int = 256
    dropout: float = 0.1
    nheads: int = 8
    num_queries: int = 30
    pre_norm: bool = False
    use_xsa: bool = False


class Agent(nn.Module):
    def __init__(self, state_dim: int, act_dim: int, args: Args):
        super().__init__()
        self.device = 'cuda'
        self.state_dim  = state_dim
        self.act_dim    = act_dim
        self.normalize  = T.Normalize(mean=[0.485, 0.456, 0.406],
                                      std=[0.229, 0.224, 0.225])
        self.include_rgb = args.include_rgb

        # CNN backbone — None for state-only mode (DETRVAE handles both paths)
        backbones = [build_backbone(args)] if args.include_rgb else None

        # CVAE decoder
        transformer = build_transformer(args)

        # CVAE encoder
        encoder = build_encoder(args)

        # ACT ( CVAE encoder + (CNN backbones + CVAE decoder) )
        self.model = DETRVAE(
            backbones,
            transformer,
            encoder,
            state_dim=state_dim,
            action_dim=act_dim,
            num_queries=args.num_queries,
        )



    def _preprocess_rgb(self, obs: dict) -> None:
        if self.include_rgb and 'rgb' in obs:
            obs['rgb'] = obs['rgb'].float() / 255.0
            # obs['rgb']: (B, num_cams, 3, 224, 224)
            B, N, C, H, W = obs['rgb'].shape
            obs['rgb'] = self.normalize(obs['rgb'].view(B * N, C, H, W)).view(B, N, C, H, W)

    def _model_input(self, obs: dict):
        # DETRVAE state-only path expects the state tensor directly, not a dict
        return obs if self.include_rgb else obs['state']

    def get_action(self, obs: dict) -> torch.Tensor:
        self._preprocess_rgb(obs)
        a_hat, _ = self.model(self._model_input(obs))
        return a_hat



class AlgSolution:

    # Slice into proprio for joint positions (relative to default).
    _QPOS_SLICE = slice(0, 8)
    _QVEL_SLICE = slice(8, 16)
    _RGB_CHANNELS = 3
    _CONCAT_IMAGE_CHANNELS = 8

    def __init__(self):
        self.device = 'cuda'
        # Default to the submission-layout policy file, but allow local eval to
        # point at a checkpoint without copying 100MB+ files around.
        policy_path = os.environ.get("ATEC_ACT_POLICY_PATH", current_path + '/policy_act.pt')
        ckpt = torch.load(policy_path, map_location=self.device)
        norm_stats = ckpt["norm_stats"]
        state_dim  = norm_stats["state_mean"].shape[-1]
        act_dim    = norm_stats["action_mean"].shape[-1]
        weight_key = "ema_agent"# if use_ema and "ema_agent" in ckpt else "agent"

        train_args = Args()
        model_args = ckpt.get("model_args", {})
        for key in (
            "enc_layers",
            "dec_layers",
            "dim_feedforward",
            "hidden_dim",
            "dropout",
            "nheads",
            "num_queries",
            "pre_norm",
            "use_xsa",
        ):
            if key in model_args:
                setattr(train_args, key, model_args[key])
        train_args.include_rgb = model_args.get(
            "include_rgb",
            any("backbone" in k for k in ckpt[weight_key].keys()),
        )

        self.agent = Agent(state_dim, act_dim, train_args).to(self.device)
        self.agent.load_state_dict(ckpt[weight_key])
        self.agent.eval()

        self.num_queries  = train_args.num_queries
        self.temporal_agg = os.environ.get("ATEC_ACT_TEMPORAL_AGG", "1").lower() not in ("0", "false", "no")
        self._k           = float(os.environ.get("ATEC_ACT_TEMPORAL_K", "0.01"))
        self._prefer_new_actions = os.environ.get("ATEC_ACT_PREFER_NEW", "0").lower() in ("1", "true", "yes")

        self.state_mean = norm_stats["state_mean"].to(self.device)   # (1, state_dim)
        self.state_std  = norm_stats["state_std"].to(self.device)    # (1, state_dim)
        self.act_mean   = norm_stats["action_mean"].to(self.device)  # (1, act_dim)
        self.act_std    = norm_stats["action_std"].to(self.device)   # (1, act_dim)

        self.default_joint_pos = torch.tensor(
            [[0.0, 1.2, -1.5, 0.0, 1.2, 0.0, 0.035, -0.035]],
            dtype=torch.float32,
            device=self.device,
        )
        # Per-episode state
        self._ts: int = 0
        self._action_history: deque = deque(maxlen=self.num_queries)
        self._last_action_seq: torch.Tensor | None = None


        startup_zero_steps = 25
        home_qpos_tolerance = 0.10
        home_hold_steps = 5

        self.teleop_home_joint_pos = torch.tensor(
            [[-0.000033, 0.924525, -1.514983, 0.000011, 1.219900, -0.000033, 0.035000, -0.035000]],
            dtype=torch.float32,
            device=self.device,
        )

        self._startup_zero_steps = max(0, int(startup_zero_steps))
        self._home_qpos_tolerance = float(home_qpos_tolerance)
        self._home_hold_steps = max(0, int(home_hold_steps))
        self._home_action = torch.clamp(
            (self.teleop_home_joint_pos - self.default_joint_pos) / 0.5,
            -1.0,
            1.0,
        )

        self._startup_step = 0
        self._home_stable_steps = 0
        self._home_done = False

    def reset_episode(self):
        self._ts = 0
        self._action_history.clear()
        self._last_action_seq = None
        self._startup_step = 0
        self._home_stable_steps = 0
        self._home_done = False

    def get_action_spec(self):
        # Use the official default Task-E Piper action configuration.
        return None


    def _compute_home_action(self, proprio):
        joint_pos_rel = proprio[:, self._QPOS_SLICE]
        qpos = joint_pos_rel + self.default_joint_pos
        qerr = self.teleop_home_joint_pos - qpos

        within_tolerance = torch.all(torch.abs(qerr) <= self._home_qpos_tolerance, dim=1)
        self._home_stable_steps = self._home_stable_steps + 1 if bool(torch.all(within_tolerance)) else 0

        # Env action is a relative joint-position target, not velocity/torque.
        # Keep commanding the absolute teleop-home target until the ACT rollout starts.
        action = self._home_action.repeat(proprio.shape[0], 1)
        home_reached = self._home_stable_steps >= self._home_hold_steps
        return action, home_reached


    def predicts(self, obs, current_score):
        if not isinstance(obs, dict) or "proprio" not in obs:
            raise ValueError("Expected obs dict with 'proprio' key.")

        proprio = obs["proprio"].to(self.device)              # (num_envs, 24)

        # Stage 1: output zero actions for the first few steps.
        if self._startup_step < self._startup_zero_steps:
            self._startup_step += 1
            return {'action': torch.zeros((proprio.shape[0], self.agent.act_dim)).numpy().tolist(), 'giveup': False}

        # Stage 2: move to teleop_home using only observations.
        if not self._home_done:
            home_action, home_reached = self._compute_home_action(proprio)
            if home_reached:
                self._home_done = True
                self._ts = 0
                self._action_history.clear()
                self._last_action_seq = None
            return {'action': home_action.cpu().numpy().tolist(), 'giveup': False}

        # Recover absolute joint positions from relative obs.
        joint_pos_rel = proprio[:, self._QPOS_SLICE]          # (num_envs, 8)
        qpos  = joint_pos_rel + self.default_joint_pos        # (num_envs, 8)
        state = (qpos - self.state_mean) / self.state_std     # (num_envs, 8)
        model_obs = {"state": state}

        if self.agent.include_rgb:
            rgb = obs["image"]["video_rgb"].to(self.device)
            if rgb.shape[1] == 4:
                rgb = rgb[:, :3]                               # drop alpha if RGBA/NCHW
            if rgb.ndim == 4 and rgb.shape[-1] == 4:
                rgb = rgb[..., :3]                             # drop alpha if RGBA/NHWC
            if rgb.dtype != torch.uint8:
                rgb = (rgb.float() * 255.0).clamp(0, 255).to(torch.uint8)
            if rgb.ndim == 4 and rgb.shape[1] in (3, 4):
                pass
            else:
                rgb = rgb.permute(0, 3, 1, 2)
            if rgb.shape[-2:] != (224, 224):
                rgb = TF.resize(rgb, [224, 224],
                                interpolation=TF.InterpolationMode.BILINEAR,
                                antialias=True)
            model_obs["rgb"] = rgb.unsqueeze(1)                # (num_envs, 1, 3, 224, 224) uint8

        ts = self._ts
        query_frequency = 1 if self.temporal_agg else self.num_queries

        if ts % query_frequency == 0:
            with torch.no_grad():
                action_seq = self.agent.get_action(model_obs)  # (num_envs, num_queries, act_dim)
            if self.temporal_agg:
                self._action_history.append(action_seq)
            else:
                self._last_action_seq = action_seq

        if self.temporal_agg:
            n = len(self._action_history)
            # deque[i=0] = oldest (added n-1 steps ago); for current step its offset = n-1-i
            actions_for_curr = torch.stack(
                [seq[:, n - 1 - i, :] for i, seq in enumerate(self._action_history)],
                dim=1,
            )  # (num_envs, n, act_dim)

            # Default preserves the original convention.  ATECs long-horizon
            # rollout can also be evaluated with newer predictions weighted
            # higher via ATEC_ACT_PREFER_NEW=1.
            order = torch.arange(n, device=self.device)
            if self._prefer_new_actions:
                order = torch.flip(order, dims=[0])
            exp_weights = torch.exp(-self._k * order)
            exp_weights = (exp_weights / exp_weights.sum()).unsqueeze(0).unsqueeze(-1)
            raw_action = (actions_for_curr * exp_weights).sum(dim=1)   # (num_envs, act_dim)
        else:
            raw_action = self._last_action_seq[:, ts % query_frequency]  # (num_envs, act_dim)

        # Denormalise → env action format
        action = raw_action * self.act_std + self.act_mean
        self._ts += 1
        return {'action': action.tolist(), 'giveup': False}

	import torch
	import torch.nn as nn
	from collections import deque
	import torchvision.transforms.functional as TF
	import torchvision.transforms as T
	from dataclasses import dataclass
	import sys
	import os

	current_path = os.path.dirname(os.path.abspath(__file__))
	#sys.path.insert(0, current_path)

	from act.detr.backbone import build_backbone
	from act.detr.transformer import build_transformer
	from act.detr.detr_vae import build_encoder, DETRVAE

	@dataclass
	class Args:
	torch_deterministic: bool = True
	"""if toggled, `torch.backends.cudnn.deterministic=False`"""
	cuda: bool = True
	"""if toggled, cuda will be enabled by default"""
	temporal_agg: bool = True
	"""if toggled, temporal ensembling will be performed at inference"""

	# Backbone
	position_embedding: str = 'sine'
	backbone: str = 'resnet18'
	lr_backbone: float = 1e-5
	masks: bool = False
	dilation: bool = False
	include_depth: bool = False
	"""always False — depth not collected; kept for backbone API compatibility"""
	include_rgb: bool = True
	"""use RGB images as input (requires --save_images during collection)"""

	# Transformer
	enc_layers: int = 2
	dec_layers: int = 4
	dim_feedforward: int = 512
	hidden_dim: int = 256
	dropout: float = 0.1
	nheads: int = 8
	num_queries: int = 30
	pre_norm: bool = False
	use_xsa: bool = False


	class Agent(nn.Module):
	def __init__(self, state_dim: int, act_dim: int, args: Args):
	super().__init__()
	self.device = 'cuda'
	self.state_dim = state_dim
	self.act_dim = act_dim
	self.normalize = T.Normalize(mean=[0.485, 0.456, 0.406],
	std=[0.229, 0.224, 0.225])
	self.include_rgb = args.include_rgb

	# CNN backbone — None for state-only mode (DETRVAE handles both paths)
	backbones = [build_backbone(args)] if args.include_rgb else None

	# CVAE decoder
	transformer = build_transformer(args)

	# CVAE encoder
	encoder = build_encoder(args)

	# ACT ( CVAE encoder + (CNN backbones + CVAE decoder) )
	self.model = DETRVAE(
	backbones,
	transformer,
	encoder,
	state_dim=state_dim,
	action_dim=act_dim,
	num_queries=args.num_queries,
	)



	def _preprocess_rgb(self, obs: dict) -> None:
	if self.include_rgb and 'rgb' in obs:
	obs['rgb'] = obs['rgb'].float() / 255.0
	# obs['rgb']: (B, num_cams, 3, 224, 224)
	B, N, C, H, W = obs['rgb'].shape
	obs['rgb'] = self.normalize(obs['rgb'].view(B * N, C, H, W)).view(B, N, C, H, W)

	def _model_input(self, obs: dict):
	# DETRVAE state-only path expects the state tensor directly, not a dict
	return obs if self.include_rgb else obs['state']

	def get_action(self, obs: dict) -> torch.Tensor:
	self._preprocess_rgb(obs)
	a_hat, _ = self.model(self._model_input(obs))
	return a_hat



	class AlgSolution:

	# Slice into proprio for joint positions (relative to default).
	_QPOS_SLICE = slice(0, 8)
	_QVEL_SLICE = slice(8, 16)
	_RGB_CHANNELS = 3
	_CONCAT_IMAGE_CHANNELS = 8

	def __init__(self):
	self.device = 'cuda'
	# Default to the submission-layout policy file, but allow local eval to
	# point at a checkpoint without copying 100MB+ files around.
	policy_path = os.environ.get("ATEC_ACT_POLICY_PATH", current_path + '/policy_act.pt')
	ckpt = torch.load(policy_path, map_location=self.device)
	norm_stats = ckpt["norm_stats"]
	state_dim = norm_stats["state_mean"].shape[-1]
	act_dim = norm_stats["action_mean"].shape[-1]
	weight_key = "ema_agent"# if use_ema and "ema_agent" in ckpt else "agent"

	train_args = Args()
	model_args = ckpt.get("model_args", {})
	for key in (
	"enc_layers",
	"dec_layers",
	"dim_feedforward",
	"hidden_dim",
	"dropout",
	"nheads",
	"num_queries",
	"pre_norm",
	"use_xsa",
	):
	if key in model_args:
	setattr(train_args, key, model_args[key])
	train_args.include_rgb = model_args.get(
	"include_rgb",
	any("backbone" in k for k in ckpt[weight_key].keys()),
	)

	self.agent = Agent(state_dim, act_dim, train_args).to(self.device)
	self.agent.load_state_dict(ckpt[weight_key])
	self.agent.eval()

	self.num_queries = train_args.num_queries
	self.temporal_agg = os.environ.get("ATEC_ACT_TEMPORAL_AGG", "1").lower() not in ("0", "false", "no")
	self._k = float(os.environ.get("ATEC_ACT_TEMPORAL_K", "0.01"))
	self._prefer_new_actions = os.environ.get("ATEC_ACT_PREFER_NEW", "0").lower() in ("1", "true", "yes")

	self.state_mean = norm_stats["state_mean"].to(self.device) # (1, state_dim)
	self.state_std = norm_stats["state_std"].to(self.device) # (1, state_dim)
	self.act_mean = norm_stats["action_mean"].to(self.device) # (1, act_dim)
	self.act_std = norm_stats["action_std"].to(self.device) # (1, act_dim)

	self.default_joint_pos = torch.tensor(
	[[0.0, 1.2, -1.5, 0.0, 1.2, 0.0, 0.035, -0.035]],
	dtype=torch.float32,
	device=self.device,
	)
	# Per-episode state
	self._ts: int = 0
	self._action_history: deque = deque(maxlen=self.num_queries)
	self._last_action_seq: torch.Tensor \| None = None


	startup_zero_steps = 25
	home_qpos_tolerance = 0.10
	home_hold_steps = 5

	self.teleop_home_joint_pos = torch.tensor(
	[[-0.000033, 0.924525, -1.514983, 0.000011, 1.219900, -0.000033, 0.035000, -0.035000]],
	dtype=torch.float32,
	device=self.device,
	)

	self._startup_zero_steps = max(0, int(startup_zero_steps))
	self._home_qpos_tolerance = float(home_qpos_tolerance)
	self._home_hold_steps = max(0, int(home_hold_steps))
	self._home_action = torch.clamp(
	(self.teleop_home_joint_pos - self.default_joint_pos) / 0.5,
	-1.0,
	1.0,
	)

	self._startup_step = 0
	self._home_stable_steps = 0
	self._home_done = False

	def reset_episode(self):
	self._ts = 0
	self._action_history.clear()
	self._last_action_seq = None
	self._startup_step = 0
	self._home_stable_steps = 0
	self._home_done = False

	def get_action_spec(self):
	# Use the official default Task-E Piper action configuration.
	return None


	def _compute_home_action(self, proprio):
	joint_pos_rel = proprio[:, self._QPOS_SLICE]
	qpos = joint_pos_rel + self.default_joint_pos
	qerr = self.teleop_home_joint_pos - qpos

	within_tolerance = torch.all(torch.abs(qerr) <= self._home_qpos_tolerance, dim=1)
	self._home_stable_steps = self._home_stable_steps + 1 if bool(torch.all(within_tolerance)) else 0

	# Env action is a relative joint-position target, not velocity/torque.
	# Keep commanding the absolute teleop-home target until the ACT rollout starts.
	action = self._home_action.repeat(proprio.shape[0], 1)
	home_reached = self._home_stable_steps >= self._home_hold_steps
	return action, home_reached


	def predicts(self, obs, current_score):
	if not isinstance(obs, dict) or "proprio" not in obs:
	raise ValueError("Expected obs dict with 'proprio' key.")

	proprio = obs["proprio"].to(self.device) # (num_envs, 24)

	# Stage 1: output zero actions for the first few steps.
	if self._startup_step < self._startup_zero_steps:
	self._startup_step += 1
	return {'action': torch.zeros((proprio.shape[0], self.agent.act_dim)).numpy().tolist(), 'giveup': False}

	# Stage 2: move to teleop_home using only observations.
	if not self._home_done:
	home_action, home_reached = self._compute_home_action(proprio)
	if home_reached:
	self._home_done = True
	self._ts = 0
	self._action_history.clear()
	self._last_action_seq = None
	return {'action': home_action.cpu().numpy().tolist(), 'giveup': False}

	# Recover absolute joint positions from relative obs.
	joint_pos_rel = proprio[:, self._QPOS_SLICE] # (num_envs, 8)
	qpos = joint_pos_rel + self.default_joint_pos # (num_envs, 8)
	state = (qpos - self.state_mean) / self.state_std # (num_envs, 8)
	model_obs = {"state": state}

	if self.agent.include_rgb:
	rgb = obs["image"]["video_rgb"].to(self.device)
	if rgb.shape[1] == 4:
	rgb = rgb[:, :3] # drop alpha if RGBA/NCHW
	if rgb.ndim == 4 and rgb.shape[-1] == 4:
	rgb = rgb[..., :3] # drop alpha if RGBA/NHWC
	if rgb.dtype != torch.uint8:
	rgb = (rgb.float() * 255.0).clamp(0, 255).to(torch.uint8)
	if rgb.ndim == 4 and rgb.shape[1] in (3, 4):
	pass
	else:
	rgb = rgb.permute(0, 3, 1, 2)
	if rgb.shape[-2:] != (224, 224):
	rgb = TF.resize(rgb, [224, 224],
	interpolation=TF.InterpolationMode.BILINEAR,
	antialias=True)
	model_obs["rgb"] = rgb.unsqueeze(1) # (num_envs, 1, 3, 224, 224) uint8

	ts = self._ts
	query_frequency = 1 if self.temporal_agg else self.num_queries

	if ts % query_frequency == 0:
	with torch.no_grad():
	action_seq = self.agent.get_action(model_obs) # (num_envs, num_queries, act_dim)
	if self.temporal_agg:
	self._action_history.append(action_seq)
	else:
	self._last_action_seq = action_seq

	if self.temporal_agg:
	n = len(self._action_history)
	# deque[i=0] = oldest (added n-1 steps ago); for current step its offset = n-1-i
	actions_for_curr = torch.stack(
	[seq[:, n - 1 - i, :] for i, seq in enumerate(self._action_history)],
	dim=1,
	) # (num_envs, n, act_dim)

	# Default preserves the original convention. ATECs long-horizon
	# rollout can also be evaluated with newer predictions weighted
	# higher via ATEC_ACT_PREFER_NEW=1.
	order = torch.arange(n, device=self.device)
	if self._prefer_new_actions:
	order = torch.flip(order, dims=[0])
	exp_weights = torch.exp(-self._k * order)
	exp_weights = (exp_weights / exp_weights.sum()).unsqueeze(0).unsqueeze(-1)
	raw_action = (actions_for_curr * exp_weights).sum(dim=1) # (num_envs, act_dim)
	else:
	raw_action = self._last_action_seq[:, ts % query_frequency] # (num_envs, act_dim)

	# Denormalise → env action format
	action = raw_action * self.act_std + self.act_mean
	self._ts += 1
	return {'action': action.tolist(), 'giveup': False}