Add GR00T N1.7 training and rollout support

frontend/src/components/training/ConfigurationTab.tsx

@@ -1,5 +1,6 @@
 import React from 'react';
 import EssentialsCard from './config/EssentialsCard';
+import GrootCard from './config/GrootCard';
 import AdvancedCard from './config/AdvancedCard';
 import TargetCard from './config/TargetCard';
 import { ConfigComponentProps } from './types';
@@ -38,6 +39,7 @@ const ConfigurationTab: React.FC<ConfigurationTabProps> = ({
         datasets={datasets}
         datasetsLoading={datasetsLoading}
       />
+      <GrootCard config={config} updateConfig={updateConfig} />
       <AdvancedCard config={config} updateConfig={updateConfig} />
     </div>
   );

frontend/src/components/training/config/EssentialsCard.tsx

@@ -92,6 +92,7 @@ const EssentialsCard: React.FC<EssentialsCardProps> = ({ config, updateConfig, d
                 <SelectItem value="diffusion">Diffusion Policy</SelectItem>
                 <SelectItem value="pi0">PI0</SelectItem>
                 <SelectItem value="smolvla">SmolVLA</SelectItem>
+                <SelectItem value="groot">GR00T N1.7</SelectItem>
                 <SelectItem value="tdmpc">TD-MPC</SelectItem>
                 <SelectItem value="vqbet">VQ-BeT</SelectItem>
                 <SelectItem value="pi0_fast">PI0 Fast</SelectItem>

frontend/src/components/training/config/GrootCard.tsx

@@ -0,0 +1,178 @@
+import React, { useEffect, useRef } from 'react';
+import { Card, CardContent, CardHeader, CardTitle } from '@/components/ui/card';
+import { Input } from '@/components/ui/input';
+import { NumberInput } from '@/components/ui/number-input';
+import { Label } from '@/components/ui/label';
+import { Switch } from '@/components/ui/switch';
+import { ConfigComponentProps } from '../types';
+
+// Sensible GR00T N1.7 defaults, seeded the first time the user picks the policy
+// so the form matches the canonical fine-tuning command out of the box.
+const GROOT_DEFAULTS = {
+  policy_base_model_path: 'nvidia/GR00T-N1.7-3B',
+  policy_embodiment_tag: 'new_embodiment',
+  policy_chunk_size: 16,
+  policy_n_action_steps: 16,
+  policy_use_relative_actions: true,
+  policy_relative_exclude_joints: ['gripper'],
+  policy_use_bf16: true,
+  dataset_image_transforms_enable: true,
+} as const;
+
+const GrootCard: React.FC<ConfigComponentProps> = ({ config, updateConfig }) => {
+  const isGroot = config.policy_type === 'groot';
+  const seededDefaults = useRef(false);
+
+  // Seed defaults once when groot is selected. Each key is only written when
+  // still undefined, so the effect can't loop and never clobbers user edits.
+  useEffect(() => {
+    if (!isGroot || seededDefaults.current) return;
+    seededDefaults.current = true;
+    (Object.keys(GROOT_DEFAULTS) as (keyof typeof GROOT_DEFAULTS)[]).forEach((key) => {
+      if (config[key] === undefined) {
+        updateConfig(key, GROOT_DEFAULTS[key] as never);
+      }
+    });
+    // eslint-disable-next-line react-hooks/exhaustive-deps
+  }, [isGroot]);
+
+  if (!isGroot) return null;
+
+  const excludeJoints = (config.policy_relative_exclude_joints ?? []).join(', ');
+
+  return (
+    <Card className="bg-slate-800/50 border-slate-700 rounded-xl">
+      <CardHeader>
+        <CardTitle className="text-white">GR00T N1.7</CardTitle>
+      </CardHeader>
+      <CardContent className="space-y-6">
+        <div>
+          <Label htmlFor="policy_base_model_path" className="text-slate-300">
+            Base Model Path
+          </Label>
+          <Input
+            id="policy_base_model_path"
+            value={config.policy_base_model_path ?? ''}
+            onChange={(e) =>
+              updateConfig('policy_base_model_path', e.target.value || undefined)
+            }
+            placeholder="nvidia/GR00T-N1.7-3B"
+            className="bg-slate-900 border-slate-600 text-white rounded-lg"
+          />
+          <p className="text-xs text-slate-500 mt-1">
+            HuggingFace repo of the pretrained GR00T backbone to fine-tune.
+          </p>
+        </div>
+
+        <div className="grid grid-cols-1 md:grid-cols-2 gap-4">
+          <div>
+            <Label htmlFor="policy_embodiment_tag" className="text-slate-300">
+              Embodiment Tag
+            </Label>
+            <Input
+              id="policy_embodiment_tag"
+              value={config.policy_embodiment_tag ?? ''}
+              onChange={(e) =>
+                updateConfig('policy_embodiment_tag', e.target.value || undefined)
+              }
+              placeholder="new_embodiment"
+              className="bg-slate-900 border-slate-600 text-white rounded-lg"
+            />
+          </div>
+
+          <div>
+            <Label htmlFor="policy_relative_exclude_joints" className="text-slate-300">
+              Relative Exclude Joints
+            </Label>
+            <Input
+              id="policy_relative_exclude_joints"
+              value={excludeJoints}
+              onChange={(e) =>
+                updateConfig(
+                  'policy_relative_exclude_joints',
+                  e.target.value
+                    .split(',')
+                    .map((s) => s.trim())
+                    .filter(Boolean),
+                )
+              }
+              placeholder="gripper"
+              className="bg-slate-900 border-slate-600 text-white rounded-lg"
+            />
+            <p className="text-xs text-slate-500 mt-1">
+              Comma-separated joints kept absolute when relative actions are on.
+            </p>
+          </div>
+
+          <div>
+            <Label htmlFor="policy_chunk_size" className="text-slate-300">
+              Chunk Size
+            </Label>
+            <NumberInput
+              id="policy_chunk_size"
+              value={config.policy_chunk_size}
+              onChange={(v) => updateConfig('policy_chunk_size', v)}
+              className="bg-slate-900 border-slate-600 text-white rounded-lg"
+            />
+          </div>
+
+          <div>
+            <Label htmlFor="policy_n_action_steps" className="text-slate-300">
+              Action Steps
+            </Label>
+            <NumberInput
+              id="policy_n_action_steps"
+              value={config.policy_n_action_steps}
+              onChange={(v) => updateConfig('policy_n_action_steps', v)}
+              className="bg-slate-900 border-slate-600 text-white rounded-lg"
+            />
+          </div>
+        </div>
+
+        <div className="space-y-3">
+          <div className="flex items-center space-x-3">
+            <Switch
+              id="policy_use_relative_actions"
+              checked={config.policy_use_relative_actions ?? false}
+              onCheckedChange={(checked) =>
+                updateConfig('policy_use_relative_actions', checked)
+              }
+              className="data-[state=checked]:bg-green-500"
+            />
+            <Label htmlFor="policy_use_relative_actions" className="text-slate-300">
+              Use Relative Actions
+            </Label>
+          </div>
+
+          <div className="flex items-center space-x-3">
+            <Switch
+              id="policy_use_bf16"
+              checked={config.policy_use_bf16 ?? false}
+              onCheckedChange={(checked) => updateConfig('policy_use_bf16', checked)}
+              className="data-[state=checked]:bg-green-500"
+            />
+            <Label htmlFor="policy_use_bf16" className="text-slate-300">
+              Use bfloat16
+            </Label>
+          </div>
+
+          <div className="flex items-center space-x-3">
+            <Switch
+              id="dataset_image_transforms_enable"
+              checked={config.dataset_image_transforms_enable ?? false}
+              onCheckedChange={(checked) =>
+                updateConfig('dataset_image_transforms_enable', checked)
+              }
+              className="data-[state=checked]:bg-green-500"
+            />
+            <Label htmlFor="dataset_image_transforms_enable" className="text-slate-300">
+              Enable Image Augmentations
+            </Label>
+          </div>
+        </div>
+      </CardContent>
+    </Card>
+  );
+};
+
+export default GrootCard;

frontend/src/components/training/types.ts

@@ -41,6 +41,16 @@ export interface TrainingConfig {
 
   // Advanced configuration
   use_policy_training_preset: boolean;
+
+  // GR00T-specific configuration (only used when policy_type === "groot").
+  dataset_image_transforms_enable?: boolean;
+  policy_base_model_path?: string;
+  policy_embodiment_tag?: string;
+  policy_chunk_size?: number;
+  policy_n_action_steps?: number;
+  policy_use_relative_actions?: boolean;
+  policy_relative_exclude_joints?: string[];
+  policy_use_bf16?: boolean;
 }
 
 export interface TrainingStatus {

frontend/src/lib/jobsApi.ts

@@ -49,6 +49,16 @@ export interface TrainingRequest {
   optimizer_weight_decay?: number;
   optimizer_grad_clip_norm?: number;
   use_policy_training_preset: boolean;
+  // GR00T-specific (only sent when policy_type === "groot"); backend ignores
+  // the policy_* ones for other policies.
+  dataset_image_transforms_enable?: boolean;
+  policy_base_model_path?: string;
+  policy_embodiment_tag?: string;
+  policy_chunk_size?: number;
+  policy_n_action_steps?: number;
+  policy_use_relative_actions?: boolean;
+  policy_relative_exclude_joints?: string[];
+  policy_use_bf16?: boolean;
   // Optional target for runner dispatch; omitted ⇒ local.
   target?: { runner: "local" | "hf_cloud"; flavor?: string };
 }

frontend/src/pages/Training.tsx

@@ -93,6 +93,20 @@ function configToRequest(c: TrainingConfig): TrainingRequest {
     optimizer_weight_decay: c.optimizer_weight_decay,
     optimizer_grad_clip_norm: c.optimizer_grad_clip_norm,
     use_policy_training_preset: c.use_policy_training_preset,
+    // GR00T-specific. Only forward the policy_* fields for groot so other
+    // policies never receive flags their config doesn't define.
+    dataset_image_transforms_enable: c.dataset_image_transforms_enable,
+    ...(c.policy_type === "groot"
+      ? {
+          policy_base_model_path: c.policy_base_model_path,
+          policy_embodiment_tag: c.policy_embodiment_tag,
+          policy_chunk_size: c.policy_chunk_size,
+          policy_n_action_steps: c.policy_n_action_steps,
+          policy_use_relative_actions: c.policy_use_relative_actions,
+          policy_relative_exclude_joints: c.policy_relative_exclude_joints,
+          policy_use_bf16: c.policy_use_bf16,
+        }
+      : {}),
   };
 }
 

lelab/rollout.py

@@ -24,6 +24,7 @@
 from __future__ import annotations
 
 import contextlib
+import json
 import logging
 import os
 import re
@@ -139,6 +140,53 @@ def _resolve_policy_path(policy_ref: str) -> str:
     raise ValueError(f"Unrecognised policy ref: {policy_ref!r}")
 
 
+def _read_policy_config(policy_path: str) -> dict[str, Any]:
+    """Load pretrained_model/config.json if present."""
+    config_path = Path(policy_path) / "config.json"
+    if not config_path.is_file():
+        return {}
+    try:
+        with open(config_path) as fh:
+            data = json.load(fh)
+        return data if isinstance(data, dict) else {}
+    except Exception as exc:
+        logger.warning("Failed to read policy config at %s: %s", config_path, exc)
+        return {}
+
+
+def _rollout_inference_args(policy_path: str) -> list[str]:
+    """Return extra lerobot-rollout flags for policies that reject sync inference.
+
+    GR00T and other relative-action policies decode action chunks against the
+    observation state at inference time. The default sync backend calls
+    ``select_action`` per tick and can re-decode cached relative actions
+    against newer states, so lerobot requires the RTC/chunked rollout path
+    instead.
+
+    GR00T also needs tuned RTC queue settings: the lerobot default
+    ``queue_threshold=30`` makes the background thread replan while a 16-step
+    chunk is still playing, which replaces the queue mid-motion and feels
+    like one action then a full recalculation."""
+    cfg = _read_policy_config(policy_path)
+    policy_type = cfg.get("type")
+    needs_rtc = bool(cfg.get("use_relative_actions")) or policy_type == "groot"
+    if not needs_rtc:
+        return []
+
+    args = ["--inference.type=rtc"]
+
+    n_action_steps = cfg.get("n_action_steps")
+    if isinstance(n_action_steps, int) and n_action_steps > 0:
+        args.append(f"--inference.rtc.execution_horizon={n_action_steps}")
+    elif policy_type == "groot":
+        args.append("--inference.rtc.execution_horizon=16")
+
+    # Wait until the current chunk is consumed before replanning. GROOT docs
+    # recommend keeping this at 0 (never > 5) for stable real-robot rollout.
+    args.append("--inference.queue_threshold=0")
+    return args
+
+
 def _format_cameras_arg(cameras: dict[str, dict[str, Any]]) -> str:
     """Convert {name: {type, camera_index, width, height, fps}} into
     lerobot's CLI dict syntax. The frontend key `camera_index` is
@@ -211,6 +259,11 @@ def _friendly_hint(error_text: str | None) -> str | None:
         return "A camera doesn't support the configured resolution — open camera settings and click Auto."
     if "permission" in low and ("port" in low or "com" in low):
         return "Couldn't open the serial port — close anything else using it, or run `lelab --stop`."
+    if "relative-action" in low or "relative chunk actions" in low:
+        return (
+            "This policy was trained with relative actions and needs RTC (chunked) inference. "
+            "Update lelab and try again — recent versions enable that automatically."
+        )
     return None
 
 
@@ -286,6 +339,7 @@ def handle_start_inference(request: InferenceRequest) -> dict[str, Any]:
             f"--robot.id={follower_id}",
             f"--task={request.task}",
             f"--duration={request.duration_s}",
+            *_rollout_inference_args(policy_path),
         ]
         if request.cameras:
             cmd.append(f"--robot.cameras={_format_cameras_arg(request.cameras)}")